{
 "generated_from": "data/llm-bench/results/",
 "records": [
  {
   "id": "dgx1-1gpu-deepseek-v4-1-flash-release-fp8-mxfp4",
   "model": "DeepSeek-V4.1-Flash",
   "model_slug": "deepseek-v4-1-flash",
   "hardware": "1x GB300 (DGX Station GB300, dgx1)",
   "hardware_family": "dgx-station",
   "units": 1,
   "engine": "vLLM (patched, Engram-on-disk)",
   "quant": "FP8 + MXFP4 (official release checkpoint)",
   "measured": "2026-09-25",
   "metrics": {
    "decode_1u_prose": {
     "value": 70,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-25-dgx1-best-model-eval.md",
     "quote": "70 tok/s single and 276 at 32"
    },
    "agg_decode_32u": {
     "value": 276,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-25-dgx1-best-model-eval.md",
     "quote": "70 tok/s single and 276 at 32"
    },
    "x_cmmc_closedbook_of_35": {
     "value": 31,
     "unit": "count",
     "label": "Closed-book CMMC items correct of 35 (first of three runs)",
     "source": "fleet-bench/2026-09-25-dgx1-best-model-eval.md",
     "quote": "scored CMMC 31, 33, 32"
    }
   }
  },
  {
   "id": "dgx1-1gpu-deepseek-v4-1-flash-v20-peer-tier",
   "model": "DeepSeek-V4.1-Flash",
   "model_slug": "deepseek-v4-1-flash",
   "hardware": "1x GB300 (dgx1) + RTX PRO 6000 cold-expert peer tier",
   "hardware_family": "dgx-station",
   "units": 2,
   "engine": "vLLM",
   "quant": "MXFP4 experts (official checkpoint, Al-ENGR v20 pin-hot-experts lane) + DSpark k-schedule",
   "measured": "2026-10-20",
   "metrics": {
    "decode_1u_prose": {
     "value": 180,
     "unit": "tok/s",
     "source": "fleet-bench/ds41-v20-dgx1-2026-10-20/RESULTS.md",
     "quote": "| 1 | 158.3 | **172.1** | +8.7% | 164 -> **180** | 0.07 s | 217 |",
     "column": "per stream b -> c",
     "part": 1
    },
    "peak_agg_decode": {
     "value": 1054.1,
     "unit": "tok/s",
     "source": "fleet-bench/ds41-v20-dgx1-2026-10-20/RESULTS.md",
     "quote": "| 24 | 729.2 | **1,054.1** | +44.6% | 45 -> 53 | 0.86 s | - |",
     "column": "run c agg",
     "part": 0
    },
    "context_verified": {
     "value": 973352,
     "unit": "tokens",
     "source": "fleet-bench/ds41-v20-dgx1-2026-10-20/RESULTS.md",
     "quote": "| **Needle 1M** | **pass**, 973,352 prompt tokens, 84 s | first 1M needle on this lane |",
     "column": "Run c",
     "part": 0
    },
    "toolcall_pass": {
     "value": 37,
     "unit": "count",
     "source": "fleet-bench/ds41-v20-dgx1-2026-10-20/RESULTS.md",
     "quote": "| Tools gate (38) | **37/38 in 18 s** | run b 37/38 in 36 s; Muse 38/38 in 277 s |",
     "column": "Run c",
     "part": 0
    },
    "cmmc_rag_accuracy_pct": {
     "value": 90,
     "unit": "%",
     "source": "fleet-bench/ds41-v20-dgx1-2026-10-20/RESULTS.md",
     "quote": "acc 91 / 90 / 90, strict 88 / 89 / 88"
    },
    "cmmc_rag_strict_pct": {
     "value": 88,
     "unit": "%",
     "source": "fleet-bench/ds41-v20-dgx1-2026-10-20/RESULTS.md",
     "quote": "acc 91 / 90 / 90, strict 88 / 89 / 88"
    },
    "x_humaneval_pct": {
     "value": 98.8,
     "unit": "%",
     "label": "HumanEval pass@1 (vendor sampling)",
     "source": "fleet-bench/ds41-v20-dgx1-2026-10-20/RESULTS.md",
     "quote": "**HumanEval 162/164 (98.8%)**"
    },
    "x_agg_decode_16u": {
     "value": 878.5,
     "unit": "tok/s",
     "label": "Aggregate decode at 16 streams (run c, peer tier on)",
     "source": "fleet-bench/ds41-v20-dgx1-2026-10-20/RESULTS.md",
     "quote": "| 16 | 754.6 | **878.5** | +16.4% | 56 -> 65 | 0.63 s | 1,154 |",
     "column": "run c agg",
     "part": 0
    },
    "x_agg_decode_24u": {
     "value": 1054.1,
     "unit": "tok/s",
     "source": "fleet-bench/ds41-v20-dgx1-2026-10-20/RESULTS.md",
     "quote": "| 24 | 729.2 | **1,054.1** | +44.6% | 45 -> 53 | 0.86 s | - |",
     "column": "run c agg",
     "part": 0,
     "label": "Aggregate decode at 24 streams (run c, peer tier on)"
    }
   },
   "notes": "Run c (v20 + our RTX PRO peer tier), the lane left running on dgx1. Without the peer tier (run b) the same lane measured 164.4 per stream at 1 stream and 754.6 aggregate at 16."
  },
  {
   "id": "dgx1-1gpu-deepseek-v4-flash-0731-fp8",
   "model": "DeepSeek-V4-Flash-0731",
   "model_slug": "deepseek-v4-flash",
   "hardware": "1x GB300 (DGX Station GB300, dgx1)",
   "hardware_family": "dgx-station",
   "units": 1,
   "engine": "vLLM",
   "quant": "FP8 (official 0731 checkpoint, fp8 KV)",
   "measured": "2026-10-05",
   "metrics": {
    "decode_1u_prose": {
     "value": 168,
     "unit": "tok/s",
     "source": "fleet-bench/dgx1-overnight-2026-10-05/RESULTS.md",
     "quote": "| DeepSeek-V4-Flash-0731 FP8 | **dgx1 GB300 x1** | 8/8 | **168** | **168** | **2,425** | **6,004 @128** | **39,200 / 40,494 / 36,790** | 1,008 module | 2.41 |",
     "column": "1-user prose tok/s",
     "part": null
    },
    "decode_1u_code": {
     "value": 168,
     "unit": "tok/s",
     "source": "fleet-bench/dgx1-overnight-2026-10-05/RESULTS.md",
     "quote": "| DeepSeek-V4-Flash-0731 FP8 | **dgx1 GB300 x1** | 8/8 | **168** | **168** | **2,425** | **6,004 @128** | **39,200 / 40,494 / 36,790** | 1,008 module | 2.41 |",
     "column": "1-user code tok/s",
     "part": null
    },
    "agg_decode_32u": {
     "value": 2425,
     "unit": "tok/s",
     "source": "fleet-bench/dgx1-overnight-2026-10-05/RESULTS.md",
     "quote": "| DeepSeek-V4-Flash-0731 FP8 | **dgx1 GB300 x1** | 8/8 | **168** | **168** | **2,425** | **6,004 @128** | **39,200 / 40,494 / 36,790** | 1,008 module | 2.41 |",
     "column": "Agg @32",
     "part": null
    },
    "peak_agg_decode": {
     "value": 6004,
     "unit": "tok/s",
     "source": "fleet-bench/dgx1-overnight-2026-10-05/RESULTS.md",
     "quote": "| DeepSeek-V4-Flash-0731 FP8 | **dgx1 GB300 x1** | 8/8 | **168** | **168** | **2,425** | **6,004 @128** | **39,200 / 40,494 / 36,790** | 1,008 module | 2.41 |",
     "column": "Peak agg (>=10 tok/s/user, TTFT p95<=10 s)",
     "part": 0
    },
    "prefill_8k": {
     "value": 39200,
     "unit": "tok/s",
     "source": "fleet-bench/dgx1-overnight-2026-10-05/RESULTS.md",
     "quote": "| DeepSeek-V4-Flash-0731 FP8 | **dgx1 GB300 x1** | 8/8 | **168** | **168** | **2,425** | **6,004 @128** | **39,200 / 40,494 / 36,790** | 1,008 module | 2.41 |",
     "column": "Prefill 8k / 32k / 128k tok/s",
     "part": 0
    },
    "prefill_32k": {
     "value": 40494,
     "unit": "tok/s",
     "source": "fleet-bench/dgx1-overnight-2026-10-05/RESULTS.md",
     "quote": "| DeepSeek-V4-Flash-0731 FP8 | **dgx1 GB300 x1** | 8/8 | **168** | **168** | **2,425** | **6,004 @128** | **39,200 / 40,494 / 36,790** | 1,008 module | 2.41 |",
     "column": "Prefill 8k / 32k / 128k tok/s",
     "part": 1
    },
    "prefill_128k": {
     "value": 36790,
     "unit": "tok/s",
     "source": "fleet-bench/dgx1-overnight-2026-10-05/RESULTS.md",
     "quote": "| DeepSeek-V4-Flash-0731 FP8 | **dgx1 GB300 x1** | 8/8 | **168** | **168** | **2,425** | **6,004 @128** | **39,200 / 40,494 / 36,790** | 1,008 module | 2.41 |",
     "column": "Prefill 8k / 32k / 128k tok/s",
     "part": 2
    },
    "tok_per_joule": {
     "value": 2.41,
     "unit": "tok/J",
     "source": "fleet-bench/dgx1-overnight-2026-10-05/RESULTS.md",
     "quote": "| DeepSeek-V4-Flash-0731 FP8 | **dgx1 GB300 x1** | 8/8 | **168** | **168** | **2,425** | **6,004 @128** | **39,200 / 40,494 / 36,790** | 1,008 module | 2.41 |",
     "column": "tok/J @32",
     "part": null
    },
    "power_w": {
     "value": 1008,
     "unit": "W",
     "source": "fleet-bench/dgx1-overnight-2026-10-05/RESULTS.md",
     "quote": "| DeepSeek-V4-Flash-0731 FP8 | **dgx1 GB300 x1** | 8/8 | **168** | **168** | **2,425** | **6,004 @128** | **39,200 / 40,494 / 36,790** | 1,008 module | 2.41 |",
     "column": "W @32",
     "part": null
    },
    "context_verified": {
     "value": 973405,
     "unit": "tokens",
     "source": "fleet-bench/2026-09-25-dgx1-best-model-eval.md",
     "quote": "| DeepSeek-V4-Flash-0731 (FP8, fp8 KV) | 1,048,576 | 4,881,345 | 171 | 1,987 | 3,861 (128 users, 50) | not reached at 256 (engine max-num-seqs) | 3/3 pass at 973,405 tokens, 55.6/55.5/52.8 s each |",
     "column": "needle at 90% of max (3 depths)",
     "part": 2
    },
    "x_cmmc_closedbook_of_35": {
     "value": 23,
     "unit": "count",
     "label": "Closed-book CMMC items correct of 35 (vendor sampling)",
     "source": "fleet-bench/2026-09-25-dgx1-best-model-eval.md",
     "quote": "| DeepSeek-V4-Flash-0731 FP8 | vendor | 38/38 | 15/15 | 23/35 | 9/10 | 3/3 | 26/26 | 114/127 | 162,734 | 707 |",
     "column": "cmmc",
     "part": 0
    },
    "x_humaneval_pct": {
     "value": 97.6,
     "unit": "%",
     "label": "HumanEval pass@1 (vendor sampling)",
     "source": "fleet-bench/2026-09-25-dgx1-best-model-eval.md",
     "quote": "| DeepSeek-V4-Flash-0731 FP8 | vendor | 15/15 | 1/3 | 97.6% | fail 0.79 | PASS 1.00 | PASS 0.94 | fail 0.00 | PASS 1.00 | 3/5 |",
     "column": "HumanEval pass@1",
     "part": 0
    }
   }
  },
  {
   "id": "dgx1-1gpu-gemma-4-31b-bf16",
   "model": "Gemma-4-31B-IT",
   "model_slug": "gemma-4",
   "hardware": "1x GB300 (DGX Station GB300, dgx1)",
   "hardware_family": "dgx-station",
   "units": 1,
   "engine": "vLLM",
   "quant": "BF16",
   "measured": "2026-09-23",
   "metrics": {
    "decode_1u_prose": {
     "value": 73.7,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-23-dgx1-dense-gemma4-31b-bf16.json",
     "quote": "\"measured_single_stream_tok_s\": 73.7,",
     "json_key": "measured_single_stream_tok_s"
    },
    "agg_decode_32u": {
     "value": 2111.0,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-23-dgx1-dense-gemma4-31b-bf16.json",
     "quote": "\"concurrency\": 32, \"ok\": 32, \"failed\": 0, \"wall_s\": 3.88, \"aggregate_tok_s\": 2111.0",
     "json_key": "aggregate_tok_s"
    }
   }
  },
  {
   "id": "dgx1-1gpu-gemma-4-31b-nvfp4",
   "model": "Gemma-4-31B-IT",
   "model_slug": "gemma-4",
   "hardware": "1x GB300 (DGX Station GB300, dgx1)",
   "hardware_family": "dgx-station",
   "units": 1,
   "engine": "vLLM",
   "quant": "NVFP4 (modelopt, nvidia checkpoint)",
   "measured": "2026-09-23",
   "metrics": {
    "decode_1u_prose": {
     "value": 110.0,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-23-dgx1-dense-gemma4-31b-nvfp4.json",
     "quote": "\"measured_single_stream_tok_s\": 110.0,",
     "json_key": "measured_single_stream_tok_s"
    },
    "agg_decode_32u": {
     "value": 3076.2,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-23-dgx1-dense-gemma4-31b-nvfp4.json",
     "quote": "\"concurrency\": 32, \"ok\": 32, \"failed\": 0, \"wall_s\": 2.66, \"aggregate_tok_s\": 3076.2",
     "json_key": "aggregate_tok_s"
    }
   }
  },
  {
   "id": "dgx1-1gpu-glm-5-3-flash-nvfp4-nvidia",
   "model": "GLM-5.3-Flash",
   "model_slug": "glm-5-3-flash",
   "hardware": "1x GB300 (DGX Station GB300, dgx1)",
   "hardware_family": "dgx-station",
   "units": 1,
   "engine": "vLLM",
   "quant": "NVFP4 (nvidia checkpoint, fp8 KV)",
   "measured": "2026-09-25",
   "metrics": {
    "decode_1u_prose": {
     "value": 145.9,
     "unit": "tok/s",
     "source": "fleet-bench/dgx1-glm53-flash-service.md",
     "quote": "**1 user 145.9 tok/s**"
    },
    "agg_decode_32u": {
     "value": 1464,
     "unit": "tok/s",
     "source": "fleet-bench/dgx1-glm53-flash-service.md",
     "quote": "32 users 1,464 aggregate"
    },
    "peak_agg_decode": {
     "value": 2517,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-25-dgx1-best-model-eval.md",
     "quote": "| GLM-5.3-Flash (nvidia NVFP4, fp8 KV) | 1,048,576 | 5,933,986 | 145 | 1,380 | 2,517 (256 users, 29) | not reached at 256 (engine max-num-seqs) | 3/3 pass at 968,377 tokens, 37.0/35.8/33.5 s each |",
     "column": "peak agg (users, per-user tok/s)",
     "part": 0
    },
    "context_verified": {
     "value": 968377,
     "unit": "tokens",
     "source": "fleet-bench/2026-09-25-dgx1-best-model-eval.md",
     "quote": "| GLM-5.3-Flash (nvidia NVFP4, fp8 KV) | 1,048,576 | 5,933,986 | 145 | 1,380 | 2,517 (256 users, 29) | not reached at 256 (engine max-num-seqs) | 3/3 pass at 968,377 tokens, 37.0/35.8/33.5 s each |",
     "column": "needle at 90% of max (3 depths)",
     "part": 2
    },
    "toolcall_pass": {
     "value": 38,
     "unit": "count",
     "source": "fleet-bench/2026-09-25-dgx1-best-model-eval.md",
     "quote": "| GLM-5.3-Flash NVFP4 (nvidia quant) | vendor | 38/38 | 15/15 | 26/35 | 10/10 | 2/3 | 26/26 | 117/127 | 288,590 | 720 |",
     "column": "tools",
     "part": 0
    },
    "x_cmmc_closedbook_of_35": {
     "value": 26,
     "unit": "count",
     "label": "Closed-book CMMC items correct of 35 (vendor sampling)",
     "source": "fleet-bench/2026-09-25-dgx1-best-model-eval.md",
     "quote": "| GLM-5.3-Flash NVFP4 (nvidia quant) | vendor | 38/38 | 15/15 | 26/35 | 10/10 | 2/3 | 26/26 | 117/127 | 288,590 | 720 |",
     "column": "cmmc",
     "part": 0
    }
   },
   "notes": "This is the deployed dgx1 service configuration (glm53-flash, :8888, go-live 2026-09-25 13:20 ET). MTP does not load with this checkpoint in this vLLM build."
  },
  {
   "id": "dgx1-1gpu-glm-5-3-flash-nvfp4-redhatai",
   "model": "GLM-5.3-Flash",
   "model_slug": "glm-5-3-flash",
   "hardware": "1x GB300 (DGX Station GB300, dgx1)",
   "hardware_family": "dgx-station",
   "units": 1,
   "engine": "vLLM",
   "quant": "NVFP4 (RedHatAI checkpoint, bf16 KV)",
   "measured": "2026-09-25",
   "metrics": {
    "decode_1u_prose": {
     "value": 144,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-25-dgx1-best-model-eval.md",
     "quote": "| GLM-5.3-Flash NVFP4 (RedHatAI quant) | 144 | 686 | 1,633 | 2,579 (43) | 28,610 / 33,076 / 32,511 | 1.69 / 6.96 s, 164 |",
     "column": "1 user tok/s",
     "part": null
    },
    "agg_decode_32u": {
     "value": 1633,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-25-dgx1-best-model-eval.md",
     "quote": "| GLM-5.3-Flash NVFP4 (RedHatAI quant) | 144 | 686 | 1,633 | 2,579 (43) | 28,610 / 33,076 / 32,511 | 1.69 / 6.96 s, 164 |",
     "column": "32 users agg",
     "part": null
    },
    "peak_agg_decode": {
     "value": 2579,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-25-dgx1-best-model-eval.md",
     "quote": "| GLM-5.3-Flash NVFP4 (RedHatAI quant) | 144 | 686 | 1,633 | 2,579 (43) | 28,610 / 33,076 / 32,511 | 1.69 / 6.96 s, 164 |",
     "column": "64 users agg (per user)",
     "part": 0
    },
    "prefill_8k": {
     "value": 28610,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-25-dgx1-best-model-eval.md",
     "quote": "| GLM-5.3-Flash NVFP4 (RedHatAI quant) | 144 | 686 | 1,633 | 2,579 (43) | 28,610 / 33,076 / 32,511 | 1.69 / 6.96 s, 164 |",
     "column": "prefill 8k / 32k / 131k tok/s (1 req)",
     "part": 0
    },
    "prefill_32k": {
     "value": 33076,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-25-dgx1-best-model-eval.md",
     "quote": "| GLM-5.3-Flash NVFP4 (RedHatAI quant) | 144 | 686 | 1,633 | 2,579 (43) | 28,610 / 33,076 / 32,511 | 1.69 / 6.96 s, 164 |",
     "column": "prefill 8k / 32k / 131k tok/s (1 req)",
     "part": 1
    },
    "prefill_128k": {
     "value": 32511,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-25-dgx1-best-model-eval.md",
     "quote": "| GLM-5.3-Flash NVFP4 (RedHatAI quant) | 144 | 686 | 1,633 | 2,579 (43) | 28,610 / 33,076 / 32,511 | 1.69 / 6.96 s, 164 |",
     "column": "prefill 8k / 32k / 131k tok/s (1 req)",
     "part": 2
    },
    "x_cmmc_closedbook_of_35": {
     "value": 26,
     "unit": "count",
     "label": "Closed-book CMMC items correct of 35 (vendor sampling)",
     "source": "fleet-bench/2026-09-25-dgx1-best-model-eval.md",
     "quote": "| GLM-5.3-Flash NVFP4 (RedHatAI quant) | vendor | 38/38 | 15/15 | 26/35 | 9/10 | 2/3 | 26/26 | 116/127 | 258,955 | 693 |",
     "column": "cmmc",
     "part": 0
    },
    "x_humaneval_pct": {
     "value": 97.0,
     "unit": "%",
     "label": "HumanEval pass@1 (vendor sampling)",
     "source": "fleet-bench/2026-09-25-dgx1-best-model-eval.md",
     "quote": "| GLM-5.3-Flash NVFP4 (RedHatAI quant) | vendor | 15/15 | 1/3 | 97.0% | fail 0.50 | PASS 0.83 | PASS 1.00 | fail 0.00 | fail 0.95 | 2/5 |",
     "column": "HumanEval pass@1",
     "part": 0
    }
   },
   "notes": "Eval comparison quant. Superseded for deployment by the nvidia checkpoint."
  },
  {
   "id": "dgx1-1gpu-glm-5-3-flash-nvfp4-redhatai-mtp2",
   "model": "GLM-5.3-Flash",
   "model_slug": "glm-5-3-flash",
   "hardware": "1x GB300 (DGX Station GB300, dgx1)",
   "hardware_family": "dgx-station",
   "units": 1,
   "engine": "vLLM",
   "quant": "NVFP4 (RedHatAI checkpoint) + MTP k=2",
   "measured": "2026-09-25",
   "metrics": {
    "decode_1u_prose": {
     "value": 231,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-25-dgx1-best-model-eval.md",
     "quote": "it measured 231 tok/s for one user"
    },
    "agg_decode_32u": {
     "value": 1760,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-25-dgx1-best-model-eval.md",
     "quote": "1,760 at 32 users"
    }
   },
   "notes": "MTP raises single-user speed but this quant loses measurable closed-book CMMC recall (section 4.1 of the eval)."
  },
  {
   "id": "dgx1-1gpu-glm-5-3-flash-uncensored-nvfp4-mtp3",
   "model": "GLM-5.3-Flash-UNCENSORED-NVFP4",
   "model_slug": "glm-5-3-flash-uncensored",
   "hardware": "1x GB300 (DGX Station GB300, dgx1)",
   "hardware_family": "dgx-station",
   "units": 1,
   "engine": "vLLM",
   "quant": "NVFP4 (dealignai uncensored checkpoint) + MTP k=3",
   "measured": "2026-10-03",
   "metrics": {
    "decode_1u_prose": {
     "value": 211,
     "unit": "tok/s",
     "source": "fleet-bench/dgx1-glm53-flash-service.md",
     "quote": "**115 -> 211 tok/s** (3 runs, 207-213)"
    }
   },
   "notes": "Uncensored endpoint (:8890) since 2026-10-03; internal use only, never a gateway alias."
  },
  {
   "id": "dgx1-1gpu-glm-5-3-full-nvfp4-offload",
   "model": "GLM-5.3",
   "model_slug": "glm-5-3",
   "hardware": "1x GB300 (DGX Station GB300, dgx1, routed experts offloaded to LPDDR5X)",
   "hardware_family": "dgx-station",
   "units": 1,
   "engine": "vLLM (UVA offload)",
   "quant": "NVFP4 (RadixArk, 433 GiB, experts offloaded)",
   "measured": "2026-09-24",
   "metrics": {
    "decode_1u_prose": {
     "value": 34.5,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-25-dgx1-best-model-eval.md",
     "quote": "Throughput: from the 09-24 run on this box, 34.5 tok/s for one user and a 165.5 tok/s peak at 64 users"
    },
    "peak_agg_decode": {
     "value": 165.5,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-25-dgx1-best-model-eval.md",
     "quote": "Throughput: from the 09-24 run on this box, 34.5 tok/s for one user and a 165.5 tok/s peak at 64 users"
    },
    "x_cmmc_closedbook_of_35": {
     "value": 22,
     "unit": "count",
     "label": "Closed-book CMMC items correct of 35 (vendor sampling, one run)",
     "source": "fleet-bench/2026-09-25-dgx1-best-model-eval.md",
     "quote": "| GLM-5.3 full, 1x GB300 (offload, MTP2) | 38/38 | 15/15 | 22/35 | 10/10 | 2/3 | 26/26 | 113/127 | 88.9% (0% halluc.) | 0.917 / 0.083, 9.7 turns | 5,348 s |",
     "column": "CMMC",
     "part": 0
    }
   },
   "notes": "MTP k=2 with the offload, --cpu-offload-gb 235. Not recommended: about 30x slower at peak than GLM-5.3-Flash on the same box."
  },
  {
   "id": "dgx1-1gpu-gpt-oss-120b-mxfp4",
   "model": "gpt-oss-120b",
   "model_slug": "gpt-oss-120b",
   "hardware": "1x GB300 (DGX Station GB300, dgx1)",
   "hardware_family": "dgx-station",
   "units": 1,
   "engine": "vLLM",
   "quant": "MXFP4 (as shipped)",
   "measured": "2026-09-23",
   "metrics": {
    "decode_1u_prose": {
     "value": 390.7,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-23-dgx1-moe-gpt-oss-120b-mxfp4.json",
     "quote": "\"measured_single_stream_tok_s\": 390.7,",
     "json_key": "measured_single_stream_tok_s"
    },
    "agg_decode_32u": {
     "value": 4924.2,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-23-dgx1-moe-gpt-oss-120b-mxfp4.json",
     "quote": "\"concurrency\": 32, \"ok\": 32, \"failed\": 0, \"wall_s\": 1.66, \"aggregate_tok_s\": 4924.2",
     "json_key": "aggregate_tok_s"
    }
   },
   "notes": "Measured with the reasoning-aware harness variant; thinking cannot be disabled on gpt-oss, so not comparable with the thinking-off rows."
  },
  {
   "id": "dgx1-1gpu-gpt-oss-20b-mxfp4",
   "model": "gpt-oss-20b",
   "model_slug": "gpt-oss-20b",
   "hardware": "1x GB300 (DGX Station GB300, dgx1)",
   "hardware_family": "dgx-station",
   "units": 1,
   "engine": "vLLM",
   "quant": "MXFP4 (as shipped)",
   "measured": "2026-09-23",
   "metrics": {
    "decode_1u_prose": {
     "value": 549.2,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-23-dgx1-moe-gpt-oss-20b-mxfp4.json",
     "quote": "\"measured_single_stream_tok_s\": 549.2,",
     "json_key": "measured_single_stream_tok_s"
    },
    "agg_decode_32u": {
     "value": 9593.0,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-23-dgx1-moe-gpt-oss-20b-mxfp4.json",
     "quote": "\"concurrency\": 32, \"ok\": 32, \"failed\": 0, \"wall_s\": 0.85, \"aggregate_tok_s\": 9593.0",
     "json_key": "aggregate_tok_s"
    }
   },
   "notes": "Same harmony caveat as gpt-oss-120b: measured with the reasoning-aware harness variant."
  },
  {
   "id": "dgx1-1gpu-kimi-dev-72b-bf16",
   "model": "Kimi-Dev-72B",
   "model_slug": "kimi-dev-72b",
   "hardware": "1x GB300 (DGX Station GB300, dgx1)",
   "hardware_family": "dgx-station",
   "units": 1,
   "engine": "vLLM",
   "quant": "BF16",
   "measured": "2026-09-23",
   "metrics": {
    "decode_1u_prose": {
     "value": 37.0,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-23-dgx1-dense-kimi-dev-72b-bf16.json",
     "quote": "\"measured_single_stream_tok_s\": 37.0,",
     "json_key": "measured_single_stream_tok_s"
    },
    "agg_decode_32u": {
     "value": 1112.3,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-23-dgx1-dense-kimi-dev-72b-bf16.json",
     "quote": "\"concurrency\": 32, \"ok\": 32, \"failed\": 0, \"wall_s\": 7.36, \"aggregate_tok_s\": 1112.3",
     "json_key": "aggregate_tok_s"
    }
   },
   "notes": "Largest dense model in the 09-23 matrix; per-stream rate stays flat from 1 to 48 streams so aggregate scales almost linearly."
  },
  {
   "id": "dgx1-1gpu-kolibri-1-fp8",
   "model": "Kolibri-1",
   "model_slug": "kolibri-1",
   "hardware": "1x GB300 (DGX Station GB300, dgx1)",
   "hardware_family": "dgx-station",
   "units": 1,
   "engine": "vLLM",
   "quant": "FP8 (Aleph Alpha, rel. 2026-10-03)",
   "measured": "2026-10-05",
   "metrics": {
    "decode_1u_prose": {
     "value": 214,
     "unit": "tok/s",
     "source": "fleet-bench/dgx1-overnight-2026-10-05/RESULTS.md",
     "quote": "| Kolibri-1 FP8 (Aleph Alpha, rel. 2026-10-03) | **dgx1 GB300 x1** | 8/8 | 214 | 214 | 3,826 | 10,606 @128 | 66,789 / 68,599 / 52,018 | 707 module | **5.42** |",
     "column": "1-user prose tok/s",
     "part": null
    },
    "decode_1u_code": {
     "value": 214,
     "unit": "tok/s",
     "source": "fleet-bench/dgx1-overnight-2026-10-05/RESULTS.md",
     "quote": "| Kolibri-1 FP8 (Aleph Alpha, rel. 2026-10-03) | **dgx1 GB300 x1** | 8/8 | 214 | 214 | 3,826 | 10,606 @128 | 66,789 / 68,599 / 52,018 | 707 module | **5.42** |",
     "column": "1-user code tok/s",
     "part": null
    },
    "agg_decode_32u": {
     "value": 3826,
     "unit": "tok/s",
     "source": "fleet-bench/dgx1-overnight-2026-10-05/RESULTS.md",
     "quote": "| Kolibri-1 FP8 (Aleph Alpha, rel. 2026-10-03) | **dgx1 GB300 x1** | 8/8 | 214 | 214 | 3,826 | 10,606 @128 | 66,789 / 68,599 / 52,018 | 707 module | **5.42** |",
     "column": "Agg @32",
     "part": null
    },
    "peak_agg_decode": {
     "value": 10606,
     "unit": "tok/s",
     "source": "fleet-bench/dgx1-overnight-2026-10-05/RESULTS.md",
     "quote": "| Kolibri-1 FP8 (Aleph Alpha, rel. 2026-10-03) | **dgx1 GB300 x1** | 8/8 | 214 | 214 | 3,826 | 10,606 @128 | 66,789 / 68,599 / 52,018 | 707 module | **5.42** |",
     "column": "Peak agg (>=10 tok/s/user, TTFT p95<=10 s)",
     "part": 0
    },
    "prefill_8k": {
     "value": 66789,
     "unit": "tok/s",
     "source": "fleet-bench/dgx1-overnight-2026-10-05/RESULTS.md",
     "quote": "| Kolibri-1 FP8 (Aleph Alpha, rel. 2026-10-03) | **dgx1 GB300 x1** | 8/8 | 214 | 214 | 3,826 | 10,606 @128 | 66,789 / 68,599 / 52,018 | 707 module | **5.42** |",
     "column": "Prefill 8k / 32k / 128k tok/s",
     "part": 0
    },
    "prefill_32k": {
     "value": 68599,
     "unit": "tok/s",
     "source": "fleet-bench/dgx1-overnight-2026-10-05/RESULTS.md",
     "quote": "| Kolibri-1 FP8 (Aleph Alpha, rel. 2026-10-03) | **dgx1 GB300 x1** | 8/8 | 214 | 214 | 3,826 | 10,606 @128 | 66,789 / 68,599 / 52,018 | 707 module | **5.42** |",
     "column": "Prefill 8k / 32k / 128k tok/s",
     "part": 1
    },
    "prefill_128k": {
     "value": 52018,
     "unit": "tok/s",
     "source": "fleet-bench/dgx1-overnight-2026-10-05/RESULTS.md",
     "quote": "| Kolibri-1 FP8 (Aleph Alpha, rel. 2026-10-03) | **dgx1 GB300 x1** | 8/8 | 214 | 214 | 3,826 | 10,606 @128 | 66,789 / 68,599 / 52,018 | 707 module | **5.42** |",
     "column": "Prefill 8k / 32k / 128k tok/s",
     "part": 2
    },
    "tok_per_joule": {
     "value": 5.42,
     "unit": "tok/J",
     "source": "fleet-bench/dgx1-overnight-2026-10-05/RESULTS.md",
     "quote": "| Kolibri-1 FP8 (Aleph Alpha, rel. 2026-10-03) | **dgx1 GB300 x1** | 8/8 | 214 | 214 | 3,826 | 10,606 @128 | 66,789 / 68,599 / 52,018 | 707 module | **5.42** |",
     "column": "tok/J @32",
     "part": null
    },
    "power_w": {
     "value": 707,
     "unit": "W",
     "source": "fleet-bench/dgx1-overnight-2026-10-05/RESULTS.md",
     "quote": "| Kolibri-1 FP8 (Aleph Alpha, rel. 2026-10-03) | **dgx1 GB300 x1** | 8/8 | 214 | 214 | 3,826 | 10,606 @128 | 66,789 / 68,599 / 52,018 | 707 module | **5.42** |",
     "column": "W @32",
     "part": null
    }
   },
   "notes": "Speed is the best of the overnight run; its quality battery was still open when the RESULTS file was written."
  },
  {
   "id": "dgx1-1gpu-llama-3-3-70b-bf16",
   "model": "Llama-3.3-70B-Instruct",
   "model_slug": "llama-3-3-70b",
   "hardware": "1x GB300 (DGX Station GB300, dgx1)",
   "hardware_family": "dgx-station",
   "units": 1,
   "engine": "vLLM",
   "quant": "BF16 (unsloth ungated mirror)",
   "measured": "2026-09-23",
   "metrics": {
    "decode_1u_prose": {
     "value": 38.1,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-23-dgx1-dense-llama33-70b-bf16.json",
     "quote": "\"measured_single_stream_tok_s\": 38.1,",
     "json_key": "measured_single_stream_tok_s"
    },
    "agg_decode_32u": {
     "value": 1147.5,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-23-dgx1-dense-llama33-70b-bf16.json",
     "quote": "\"concurrency\": 32, \"ok\": 32, \"failed\": 0, \"wall_s\": 7.14, \"aggregate_tok_s\": 1147.5",
     "json_key": "aggregate_tok_s"
    }
   }
  },
  {
   "id": "dgx1-1gpu-llama-3-3-70b-fp8",
   "model": "Llama-3.3-70B-Instruct",
   "model_slug": "llama-3-3-70b",
   "hardware": "1x GB300 (DGX Station GB300, dgx1)",
   "hardware_family": "dgx-station",
   "units": 1,
   "engine": "vLLM",
   "quant": "FP8 e4m3 (nvidia checkpoint)",
   "measured": "2026-09-23",
   "metrics": {
    "decode_1u_prose": {
     "value": 65.3,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-23-dgx1-dense-llama33-70b-fp8.json",
     "quote": "\"measured_single_stream_tok_s\": 65.3,",
     "json_key": "measured_single_stream_tok_s"
    },
    "agg_decode_32u": {
     "value": 2093.4,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-23-dgx1-dense-llama33-70b-fp8.json",
     "quote": "\"concurrency\": 32, \"ok\": 32, \"failed\": 0, \"wall_s\": 3.91, \"aggregate_tok_s\": 2093.4",
     "json_key": "aggregate_tok_s"
    }
   }
  },
  {
   "id": "dgx1-1gpu-mimo-v2-6-flash-fp8mxfp4",
   "model": "MiMo-V2.6-Flash-RL",
   "model_slug": "mimo-v2-6-flash",
   "hardware": "1x GB300 (DGX Station GB300, dgx1)",
   "hardware_family": "dgx-station",
   "units": 1,
   "engine": "vLLM",
   "quant": "FP8 with mxfp4 store (as shipped by XiaomiMiMo, bf16 KV)",
   "measured": "2026-09-25",
   "metrics": {
    "decode_1u_prose": {
     "value": 212,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-25-dgx1-best-model-eval.md",
     "quote": "| MiMo-V2.6-Flash-RL (bf16 KV) | 1,048,576 | 2,042,766 | 212 | 1,995 | 4,941 (256 users, 33) | not reached at 256 (engine max-num-seqs) | 3/3 pass at 971,072 tokens, 127.4/126.7/123.2 s each |",
     "column": "1 user",
     "part": 0
    },
    "agg_decode_32u": {
     "value": 1995,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-25-dgx1-best-model-eval.md",
     "quote": "| MiMo-V2.6-Flash-RL (bf16 KV) | 1,048,576 | 2,042,766 | 212 | 1,995 | 4,941 (256 users, 33) | not reached at 256 (engine max-num-seqs) | 3/3 pass at 971,072 tokens, 127.4/126.7/123.2 s each |",
     "column": "32 users agg",
     "part": 0
    },
    "peak_agg_decode": {
     "value": 4941,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-25-dgx1-best-model-eval.md",
     "quote": "| MiMo-V2.6-Flash-RL (bf16 KV) | 1,048,576 | 2,042,766 | 212 | 1,995 | 4,941 (256 users, 33) | not reached at 256 (engine max-num-seqs) | 3/3 pass at 971,072 tokens, 127.4/126.7/123.2 s each |",
     "column": "peak agg (users, per-user tok/s)",
     "part": 0
    },
    "prefill_8k": {
     "value": 47439,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-25-dgx1-best-model-eval.md",
     "quote": "| MiMo-V2.6-Flash-RL | 211 | 946 | 2,340 | 3,630 (60) | 47,439 / 47,044 / 30,726 | 1.36 / 4.23 s, 239 |",
     "column": "prefill 8k / 32k / 131k tok/s (1 req)",
     "part": 0
    },
    "prefill_32k": {
     "value": 47044,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-25-dgx1-best-model-eval.md",
     "quote": "| MiMo-V2.6-Flash-RL | 211 | 946 | 2,340 | 3,630 (60) | 47,439 / 47,044 / 30,726 | 1.36 / 4.23 s, 239 |",
     "column": "prefill 8k / 32k / 131k tok/s (1 req)",
     "part": 1
    },
    "prefill_128k": {
     "value": 30726,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-25-dgx1-best-model-eval.md",
     "quote": "| MiMo-V2.6-Flash-RL | 211 | 946 | 2,340 | 3,630 (60) | 47,439 / 47,044 / 30,726 | 1.36 / 4.23 s, 239 |",
     "column": "prefill 8k / 32k / 131k tok/s (1 req)",
     "part": 2
    },
    "context_verified": {
     "value": 971072,
     "unit": "tokens",
     "source": "fleet-bench/2026-09-25-dgx1-best-model-eval.md",
     "quote": "| MiMo-V2.6-Flash-RL (bf16 KV) | 1,048,576 | 2,042,766 | 212 | 1,995 | 4,941 (256 users, 33) | not reached at 256 (engine max-num-seqs) | 3/3 pass at 971,072 tokens, 127.4/126.7/123.2 s each |",
     "column": "needle at 90% of max (3 depths)",
     "part": 2
    },
    "x_cmmc_closedbook_of_35": {
     "value": 16,
     "unit": "count",
     "label": "Closed-book CMMC items correct of 35 (vendor sampling)",
     "source": "fleet-bench/2026-09-25-dgx1-best-model-eval.md",
     "quote": "| MiMo-V2.6-Flash-RL | vendor | 38/38 | 15/15 | 16/35 | 9/10 | 3/3 | 22/26 | 103/127 | 245,185 | 484 |",
     "column": "cmmc",
     "part": 0
    },
    "x_humaneval_pct": {
     "value": 97.0,
     "unit": "%",
     "label": "HumanEval pass@1 (vendor sampling)",
     "source": "fleet-bench/2026-09-25-dgx1-best-model-eval.md",
     "quote": "| MiMo-V2.6-Flash-RL | vendor | 14/15 | 1/3 | 97.0% | fail 0.50 | PASS 0.92 | PASS 0.81 | fail 0.00 | PASS 1.00 | 3/5 |",
     "column": "HumanEval pass@1",
     "part": 0
    }
   },
   "notes": "Peak aggregate on one GB300 in the 09-25 ceiling run (4,941 at 256 users), highest of the five Flash-class candidates. MiMo refuses embedded facts as injections in its needle failures."
  },
  {
   "id": "dgx1-1gpu-qwen3-6-35b-a3b-bf16",
   "model": "Qwen3.6-35B-A3B",
   "model_slug": "qwen3-6-35b-a3b",
   "hardware": "1x GB300 (DGX Station GB300, dgx1)",
   "hardware_family": "dgx-station",
   "units": 1,
   "engine": "vLLM",
   "quant": "BF16",
   "measured": "2026-10-05",
   "metrics": {
    "decode_1u_prose": {
     "value": 335,
     "unit": "tok/s",
     "source": "fleet-bench/dgx1-overnight-2026-10-05/RESULTS.md",
     "quote": "| Qwen3.6-35B-A3B (dgx1 BF16) | **dgx1 GB300 x1** | 8/8 | **335** | **335** | **4,359** | **10,653 @128** | **39,068 / 80,337 / 65,231** | 895 module | 4.87 |",
     "column": "1-user prose tok/s",
     "part": null
    },
    "decode_1u_code": {
     "value": 335,
     "unit": "tok/s",
     "source": "fleet-bench/dgx1-overnight-2026-10-05/RESULTS.md",
     "quote": "| Qwen3.6-35B-A3B (dgx1 BF16) | **dgx1 GB300 x1** | 8/8 | **335** | **335** | **4,359** | **10,653 @128** | **39,068 / 80,337 / 65,231** | 895 module | 4.87 |",
     "column": "1-user code tok/s",
     "part": null
    },
    "agg_decode_32u": {
     "value": 4359,
     "unit": "tok/s",
     "source": "fleet-bench/dgx1-overnight-2026-10-05/RESULTS.md",
     "quote": "| Qwen3.6-35B-A3B (dgx1 BF16) | **dgx1 GB300 x1** | 8/8 | **335** | **335** | **4,359** | **10,653 @128** | **39,068 / 80,337 / 65,231** | 895 module | 4.87 |",
     "column": "Agg @32",
     "part": null
    },
    "peak_agg_decode": {
     "value": 10653,
     "unit": "tok/s",
     "source": "fleet-bench/dgx1-overnight-2026-10-05/RESULTS.md",
     "quote": "| Qwen3.6-35B-A3B (dgx1 BF16) | **dgx1 GB300 x1** | 8/8 | **335** | **335** | **4,359** | **10,653 @128** | **39,068 / 80,337 / 65,231** | 895 module | 4.87 |",
     "column": "Peak agg (>=10 tok/s/user, TTFT p95<=10 s)",
     "part": 0
    },
    "prefill_8k": {
     "value": 39068,
     "unit": "tok/s",
     "source": "fleet-bench/dgx1-overnight-2026-10-05/RESULTS.md",
     "quote": "| Qwen3.6-35B-A3B (dgx1 BF16) | **dgx1 GB300 x1** | 8/8 | **335** | **335** | **4,359** | **10,653 @128** | **39,068 / 80,337 / 65,231** | 895 module | 4.87 |",
     "column": "Prefill 8k / 32k / 128k tok/s",
     "part": 0
    },
    "prefill_32k": {
     "value": 80337,
     "unit": "tok/s",
     "source": "fleet-bench/dgx1-overnight-2026-10-05/RESULTS.md",
     "quote": "| Qwen3.6-35B-A3B (dgx1 BF16) | **dgx1 GB300 x1** | 8/8 | **335** | **335** | **4,359** | **10,653 @128** | **39,068 / 80,337 / 65,231** | 895 module | 4.87 |",
     "column": "Prefill 8k / 32k / 128k tok/s",
     "part": 1
    },
    "prefill_128k": {
     "value": 65231,
     "unit": "tok/s",
     "source": "fleet-bench/dgx1-overnight-2026-10-05/RESULTS.md",
     "quote": "| Qwen3.6-35B-A3B (dgx1 BF16) | **dgx1 GB300 x1** | 8/8 | **335** | **335** | **4,359** | **10,653 @128** | **39,068 / 80,337 / 65,231** | 895 module | 4.87 |",
     "column": "Prefill 8k / 32k / 128k tok/s",
     "part": 2
    },
    "tok_per_joule": {
     "value": 4.87,
     "unit": "tok/J",
     "source": "fleet-bench/dgx1-overnight-2026-10-05/RESULTS.md",
     "quote": "| Qwen3.6-35B-A3B (dgx1 BF16) | **dgx1 GB300 x1** | 8/8 | **335** | **335** | **4,359** | **10,653 @128** | **39,068 / 80,337 / 65,231** | 895 module | 4.87 |",
     "column": "tok/J @32",
     "part": null
    },
    "power_w": {
     "value": 895,
     "unit": "W",
     "source": "fleet-bench/dgx1-overnight-2026-10-05/RESULTS.md",
     "quote": "| Qwen3.6-35B-A3B (dgx1 BF16) | **dgx1 GB300 x1** | 8/8 | **335** | **335** | **4,359** | **10,653 @128** | **39,068 / 80,337 / 65,231** | 895 module | 4.87 |",
     "column": "W @32",
     "part": null
    }
   }
  },
  {
   "id": "dgx1-1gpu-qwen3-8-27b-bf16",
   "model": "Qwen3.8-27B",
   "model_slug": "qwen3-8-27b",
   "hardware": "1x GB300 (DGX Station GB300, dgx1)",
   "hardware_family": "dgx-station",
   "units": 1,
   "engine": "vLLM",
   "quant": "BF16 (bf16 KV)",
   "measured": "2026-09-25",
   "metrics": {
    "decode_1u_prose": {
     "value": 90,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-25-dgx1-best-model-eval.md",
     "quote": "| Qwen3.8-27B (BF16, bf16 KV) | 262,144 | 2,793,640 | 90 | 1,842 | 4,278 (256 users, 28) | not reached at 256 (engine max-num-seqs) | 3/3 pass at 243,392 tokens, 23.7/21.7/19.5 s each |",
     "column": "1 user",
     "part": 0
    },
    "agg_decode_32u": {
     "value": 1842,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-25-dgx1-best-model-eval.md",
     "quote": "| Qwen3.8-27B (BF16, bf16 KV) | 262,144 | 2,793,640 | 90 | 1,842 | 4,278 (256 users, 28) | not reached at 256 (engine max-num-seqs) | 3/3 pass at 243,392 tokens, 23.7/21.7/19.5 s each |",
     "column": "32 users agg",
     "part": 0
    },
    "peak_agg_decode": {
     "value": 4278,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-25-dgx1-best-model-eval.md",
     "quote": "| Qwen3.8-27B (BF16, bf16 KV) | 262,144 | 2,793,640 | 90 | 1,842 | 4,278 (256 users, 28) | not reached at 256 (engine max-num-seqs) | 3/3 pass at 243,392 tokens, 23.7/21.7/19.5 s each |",
     "column": "peak agg (users, per-user tok/s)",
     "part": 0
    },
    "prefill_8k": {
     "value": 24558,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-25-dgx1-best-model-eval.md",
     "quote": "| Qwen3.8-27B BF16 | 90 | 640 | 2,243 | 3,741 (62) | 24,558 / 24,215 / 18,411 | 3.96 / 9.51 s, 142 |",
     "column": "prefill 8k / 32k / 131k tok/s (1 req)",
     "part": 0
    },
    "prefill_32k": {
     "value": 24215,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-25-dgx1-best-model-eval.md",
     "quote": "| Qwen3.8-27B BF16 | 90 | 640 | 2,243 | 3,741 (62) | 24,558 / 24,215 / 18,411 | 3.96 / 9.51 s, 142 |",
     "column": "prefill 8k / 32k / 131k tok/s (1 req)",
     "part": 1
    },
    "prefill_128k": {
     "value": 18411,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-25-dgx1-best-model-eval.md",
     "quote": "| Qwen3.8-27B BF16 | 90 | 640 | 2,243 | 3,741 (62) | 24,558 / 24,215 / 18,411 | 3.96 / 9.51 s, 142 |",
     "column": "prefill 8k / 32k / 131k tok/s (1 req)",
     "part": 2
    },
    "context_verified": {
     "value": 243392,
     "unit": "tokens",
     "source": "fleet-bench/2026-09-25-dgx1-best-model-eval.md",
     "quote": "| Qwen3.8-27B (BF16, bf16 KV) | 262,144 | 2,793,640 | 90 | 1,842 | 4,278 (256 users, 28) | not reached at 256 (engine max-num-seqs) | 3/3 pass at 243,392 tokens, 23.7/21.7/19.5 s each |",
     "column": "needle at 90% of max (3 depths)",
     "part": 2
    },
    "toolcall_pass": {
     "value": 38,
     "unit": "count",
     "source": "fleet-bench/2026-09-25-dgx1-best-model-eval.md",
     "quote": "| Qwen3.8-27B BF16 | vendor | 38/38 | 15/15 | 5/35 | 10/10 | 2/3 | 26/26 | 96/127 | 309,101 | 1021 |",
     "column": "tools",
     "part": 0
    },
    "x_cmmc_closedbook_of_35": {
     "value": 5,
     "unit": "count",
     "label": "Closed-book CMMC items correct of 35 (vendor sampling)",
     "source": "fleet-bench/2026-09-25-dgx1-best-model-eval.md",
     "quote": "| Qwen3.8-27B BF16 | vendor | 38/38 | 15/15 | 5/35 | 10/10 | 2/3 | 26/26 | 96/127 | 309,101 | 1021 |",
     "column": "cmmc",
     "part": 0
    },
    "x_humaneval_pct": {
     "value": 94.5,
     "unit": "%",
     "label": "HumanEval pass@1 (vendor sampling)",
     "source": "fleet-bench/2026-09-25-dgx1-best-model-eval.md",
     "quote": "| Qwen3.8-27B BF16 | vendor | 15/15 | 1/3 | 94.5% | fail 0.57 | PASS 0.83 | fail 0.81 | fail 0.00 | fail 0.95 | 1/5 |",
     "column": "HumanEval pass@1",
     "part": 0
    }
   }
  },
  {
   "id": "dgx1-1gpu-qwen3-8-27b-fp8",
   "model": "Qwen3.8-27B",
   "model_slug": "qwen3-8-27b",
   "hardware": "1x GB300 (DGX Station GB300, dgx1)",
   "hardware_family": "dgx-station",
   "units": 1,
   "engine": "vLLM",
   "quant": "FP8 e4m3 dynamic",
   "measured": "2026-09-23",
   "metrics": {
    "decode_1u_prose": {
     "value": 111.2,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-23-dgx1-dense-qwen38-27b-fp8.json",
     "quote": "\"measured_single_stream_tok_s\": 111.2,",
     "json_key": "measured_single_stream_tok_s"
    },
    "agg_decode_32u": {
     "value": 2510.5,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-23-dgx1-dense-qwen38-27b-fp8.json",
     "quote": "\"concurrency\": 32, \"ok\": 32, \"failed\": 0, \"wall_s\": 3.26, \"aggregate_tok_s\": 2510.5",
     "json_key": "aggregate_tok_s"
    }
   }
  },
  {
   "id": "dgx1-1gpu-qwen3-8-27b-nvfp4",
   "model": "Qwen3.8-27B",
   "model_slug": "qwen3-8-27b",
   "hardware": "1x GB300 (DGX Station GB300, dgx1)",
   "hardware_family": "dgx-station",
   "units": 1,
   "engine": "vLLM",
   "quant": "NVFP4 (modelopt_mixed, nvidia checkpoint)",
   "measured": "2026-09-23",
   "metrics": {
    "decode_1u_prose": {
     "value": 162.3,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-23-dgx1-dense-qwen38-27b-nvfp4.json",
     "quote": "\"measured_single_stream_tok_s\": 162.3,",
     "json_key": "measured_single_stream_tok_s"
    },
    "agg_decode_32u": {
     "value": 3552.3,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-23-dgx1-dense-qwen38-27b-nvfp4.json",
     "quote": "\"concurrency\": 32, \"ok\": 32, \"failed\": 0, \"wall_s\": 2.31, \"aggregate_tok_s\": 3552.3",
     "json_key": "aggregate_tok_s"
    }
   },
   "notes": "Fastest single-stream of the three Qwen3.8-27B quants on the GB300."
  },
  {
   "id": "dgx1-1gpu-qwen3-8-flash-next-nvfp4",
   "model": "Qwen3.8-Flash-Next",
   "model_slug": "qwen3-8-flash-next",
   "hardware": "1x GB300 (DGX Station GB300, dgx1)",
   "hardware_family": "dgx-station",
   "units": 1,
   "engine": "vLLM",
   "quant": "NVFP4 (RadixArk checkpoint, bf16 KV)",
   "measured": "2026-10-05",
   "metrics": {
    "decode_1u_prose": {
     "value": 205,
     "unit": "tok/s",
     "source": "fleet-bench/dgx1-overnight-2026-10-05/RESULTS.md",
     "quote": "| Qwen3.8-Flash-Next NVFP4 | **dgx1 GB300 x1** | 8/8 | **205** | **205** | **3,007** | **7,040 @128** | **43,576 / 46,665 / 46,597** | 850 module | 3.54 |",
     "column": "1-user prose tok/s",
     "part": null
    },
    "decode_1u_code": {
     "value": 205,
     "unit": "tok/s",
     "source": "fleet-bench/dgx1-overnight-2026-10-05/RESULTS.md",
     "quote": "| Qwen3.8-Flash-Next NVFP4 | **dgx1 GB300 x1** | 8/8 | **205** | **205** | **3,007** | **7,040 @128** | **43,576 / 46,665 / 46,597** | 850 module | 3.54 |",
     "column": "1-user code tok/s",
     "part": null
    },
    "agg_decode_32u": {
     "value": 3007,
     "unit": "tok/s",
     "source": "fleet-bench/dgx1-overnight-2026-10-05/RESULTS.md",
     "quote": "| Qwen3.8-Flash-Next NVFP4 | **dgx1 GB300 x1** | 8/8 | **205** | **205** | **3,007** | **7,040 @128** | **43,576 / 46,665 / 46,597** | 850 module | 3.54 |",
     "column": "Agg @32",
     "part": null
    },
    "peak_agg_decode": {
     "value": 7040,
     "unit": "tok/s",
     "source": "fleet-bench/dgx1-overnight-2026-10-05/RESULTS.md",
     "quote": "| Qwen3.8-Flash-Next NVFP4 | **dgx1 GB300 x1** | 8/8 | **205** | **205** | **3,007** | **7,040 @128** | **43,576 / 46,665 / 46,597** | 850 module | 3.54 |",
     "column": "Peak agg (>=10 tok/s/user, TTFT p95<=10 s)",
     "part": 0
    },
    "prefill_8k": {
     "value": 43576,
     "unit": "tok/s",
     "source": "fleet-bench/dgx1-overnight-2026-10-05/RESULTS.md",
     "quote": "| Qwen3.8-Flash-Next NVFP4 | **dgx1 GB300 x1** | 8/8 | **205** | **205** | **3,007** | **7,040 @128** | **43,576 / 46,665 / 46,597** | 850 module | 3.54 |",
     "column": "Prefill 8k / 32k / 128k tok/s",
     "part": 0
    },
    "prefill_32k": {
     "value": 46665,
     "unit": "tok/s",
     "source": "fleet-bench/dgx1-overnight-2026-10-05/RESULTS.md",
     "quote": "| Qwen3.8-Flash-Next NVFP4 | **dgx1 GB300 x1** | 8/8 | **205** | **205** | **3,007** | **7,040 @128** | **43,576 / 46,665 / 46,597** | 850 module | 3.54 |",
     "column": "Prefill 8k / 32k / 128k tok/s",
     "part": 1
    },
    "prefill_128k": {
     "value": 46597,
     "unit": "tok/s",
     "source": "fleet-bench/dgx1-overnight-2026-10-05/RESULTS.md",
     "quote": "| Qwen3.8-Flash-Next NVFP4 | **dgx1 GB300 x1** | 8/8 | **205** | **205** | **3,007** | **7,040 @128** | **43,576 / 46,665 / 46,597** | 850 module | 3.54 |",
     "column": "Prefill 8k / 32k / 128k tok/s",
     "part": 2
    },
    "tok_per_joule": {
     "value": 3.54,
     "unit": "tok/J",
     "source": "fleet-bench/dgx1-overnight-2026-10-05/RESULTS.md",
     "quote": "| Qwen3.8-Flash-Next NVFP4 | **dgx1 GB300 x1** | 8/8 | **205** | **205** | **3,007** | **7,040 @128** | **43,576 / 46,665 / 46,597** | 850 module | 3.54 |",
     "column": "tok/J @32",
     "part": null
    },
    "power_w": {
     "value": 850,
     "unit": "W",
     "source": "fleet-bench/dgx1-overnight-2026-10-05/RESULTS.md",
     "quote": "| Qwen3.8-Flash-Next NVFP4 | **dgx1 GB300 x1** | 8/8 | **205** | **205** | **3,007** | **7,040 @128** | **43,576 / 46,665 / 46,597** | 850 module | 3.54 |",
     "column": "W @32",
     "part": null
    },
    "context_verified": {
     "value": 243392,
     "unit": "tokens",
     "source": "fleet-bench/2026-09-25-dgx1-best-model-eval.md",
     "quote": "| Qwen3.8-Flash-Next (NVFP4, bf16 KV) | 262,144 | 6,002,653 | 201 | 2,412 | 4,670 (256 users, 28) | not reached at 256 (engine max-num-seqs) | 3/3 pass at 243,392 tokens, 10.7/13.2/6.2 s each |",
     "column": "needle at 90% of max (3 depths)",
     "part": 2
    },
    "x_cmmc_closedbook_of_35": {
     "value": 11,
     "unit": "count",
     "label": "Closed-book CMMC items correct of 35 (vendor sampling)",
     "source": "fleet-bench/2026-09-25-dgx1-best-model-eval.md",
     "quote": "| Qwen3.8-Flash-Next NVFP4 | vendor | 37/38 | 15/15 | 11/35 | 9/10 | 3/3 | 24/26 | 99/127 | 312,914 | 527 |",
     "column": "cmmc",
     "part": 0
    }
   },
   "notes": "Overnight run used 2 reps per level; the file says cross-platform ratios are not cleared for the site. dgx1 power is module draw, which flatters tok/J."
  },
  {
   "id": "dgx1-gpu0-rtxpro2000-qwen3-8-27b-abliterated-iq3s",
   "model": "Huihui-Qwen3.8-27B-abliterated",
   "model_slug": "qwen3-8-27b",
   "hardware": "1x RTX PRO 2000 Blackwell 16 GB (dgx1 GPU0)",
   "hardware_family": "dgx-station",
   "units": 1,
   "engine": "llama.cpp",
   "quant": "IQ3_S GGUF with MTP head",
   "measured": "2026-09-28",
   "metrics": {
    "x_agg_decode_4u": {
     "value": 87.8,
     "unit": "tok/s",
     "label": "Aggregate decode at 4 concurrent streams (deployed config, MTP, 4 slots, 32k q8 KV)",
     "source": "fleet-bench/2026-09-28-dgx1-huihui-qwen38-abliterated.md",
     "quote": "| **Deployed: MTP, 4 slots, 32k q8 KV** | 15.4 GiB | 32-40 | **87.8** (22-28 per stream) | 600 |",
     "column": "4 concurrent aggregate",
     "part": 0
    },
    "x_prefill_tok_s": {
     "value": 600,
     "unit": "tok/s",
     "label": "Prefill tok/s (deployed config, unspecified input size)",
     "source": "fleet-bench/2026-09-28-dgx1-huihui-qwen38-abliterated.md",
     "quote": "| **Deployed: MTP, 4 slots, 32k q8 KV** | 15.4 GiB | 32-40 | **87.8** (22-28 per stream) | 600 |",
     "column": "prefill tok/s",
     "part": 0
    },
    "x_prefill_28k": {
     "value": 496,
     "unit": "tok/s",
     "label": "Prefill tok/s on a 28,305-token prompt",
     "source": "fleet-bench/2026-09-28-dgx1-huihui-qwen38-abliterated.md",
     "quote": "A 28,305-token prompt prefilled at 496 tok/s (57 s) without running out of memory."
    }
   },
   "notes": "UNCENSORED abliterated model for the security team's authorized red-team testing on our own infrastructure only; never a gateway alias. Runs on dgx1's RTX PRO 2000 GPU0, the GB300 is untouched."
  },
  {
   "id": "gb10-1u-gemma-4-26b-nvfp4",
   "model": "Gemma-4-26B-A4B",
   "model_slug": "gemma-4",
   "hardware": "1x GB10 (DGX Spark class)",
   "hardware_family": "gb10",
   "units": 1,
   "engine": "vLLM nightly",
   "quant": "NVFP4",
   "measured": "2026-10-02",
   "metrics": {
    "decode_1u_prose": {
     "value": 31.5,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| Gemma-4-26B-A4B NVFP4 (nvidia) | vLLM nightly, no drafter published | 8/8 | 31.5 / 31.4 | 552.8 | 64 (866.4) | 6,247 / 2,689 / 784 | 111 / 131 | 6.63 | 6.7 / 184 |",
     "column": "1 user prose / code",
     "part": 0
    },
    "decode_1u_code": {
     "value": 31.4,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| Gemma-4-26B-A4B NVFP4 (nvidia) | vLLM nightly, no drafter published | 8/8 | 31.5 / 31.4 | 552.8 | 64 (866.4) | 6,247 / 2,689 / 784 | 111 / 131 | 6.63 | 6.7 / 184 |",
     "column": "1 user prose / code",
     "part": 1
    },
    "agg_decode_32u": {
     "value": 552.8,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| Gemma-4-26B-A4B NVFP4 (nvidia) | vLLM nightly, no drafter published | 8/8 | 31.5 / 31.4 | 552.8 | 64 (866.4) | 6,247 / 2,689 / 784 | 111 / 131 | 6.63 | 6.7 / 184 |",
     "column": "32 users"
    },
    "ceiling_users": {
     "value": 64,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| Gemma-4-26B-A4B NVFP4 (nvidia) | vLLM nightly, no drafter published | 8/8 | 31.5 / 31.4 | 552.8 | 64 (866.4) | 6,247 / 2,689 / 784 | 111 / 131 | 6.63 | 6.7 / 184 |",
     "column": "ceiling users (agg)",
     "part": 0
    },
    "agg_decode_at_ceiling": {
     "value": 866.4,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| Gemma-4-26B-A4B NVFP4 (nvidia) | vLLM nightly, no drafter published | 8/8 | 31.5 / 31.4 | 552.8 | 64 (866.4) | 6,247 / 2,689 / 784 | 111 / 131 | 6.63 | 6.7 / 184 |",
     "column": "ceiling users (agg)",
     "part": 1
    },
    "prefill_8k": {
     "value": 6247,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| Gemma-4-26B-A4B NVFP4 (nvidia) | vLLM nightly, no drafter published | 8/8 | 31.5 / 31.4 | 552.8 | 64 (866.4) | 6,247 / 2,689 / 784 | 111 / 131 | 6.63 | 6.7 / 184 |",
     "column": "prefill 8k / 32k / 128k",
     "part": 0
    },
    "prefill_32k": {
     "value": 2689,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| Gemma-4-26B-A4B NVFP4 (nvidia) | vLLM nightly, no drafter published | 8/8 | 31.5 / 31.4 | 552.8 | 64 (866.4) | 6,247 / 2,689 / 784 | 111 / 131 | 6.63 | 6.7 / 184 |",
     "column": "prefill 8k / 32k / 128k",
     "part": 1
    },
    "prefill_128k": {
     "value": 784,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| Gemma-4-26B-A4B NVFP4 (nvidia) | vLLM nightly, no drafter published | 8/8 | 31.5 / 31.4 | 552.8 | 64 (866.4) | 6,247 / 2,689 / 784 | 111 / 131 | 6.63 | 6.7 / 184 |",
     "column": "prefill 8k / 32k / 128k",
     "part": 2
    },
    "tok_per_joule": {
     "value": 6.63,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| Gemma-4-26B-A4B NVFP4 (nvidia) | vLLM nightly, no drafter published | 8/8 | 31.5 / 31.4 | 552.8 | 64 (866.4) | 6,247 / 2,689 / 784 | 111 / 131 | 6.63 | 6.7 / 184 |",
     "column": "tok/J at ceiling"
    },
    "context_verified": {
     "value": 243271,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| Gemma-4-26B-A4B | 1 | 262,144 | 262,144 | 5,005,860 | 19.1 | 243,271: pass, 556 s | 31.4 / 23.7 at 235,329 | 520 s | 64 |",
     "column": "needle at 90% (prompt tokens, result, request time)",
     "part": 0
    },
    "cmmc_rag_strict_pct": {
     "value": 62,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| Gemma-4-26B-A4B (1) | 59 / 62 / 55 | 77-80 | 2-4 | 9/10 | 8-9 | 370/371 | 9.3-9.9 | 218-230 | 83 | 79 | 33 |",
     "column": "agentic strict /93 (seeds)",
     "part": 1
    }
   },
   "notes": "Best one-unit quality among the MoE models (tools 37/38, CMMC 13-14 of 35, long context 24/26) and it holds the most users at the floor, but single-user speed is only about 31 tok/s (no drafter) and 128k prefill is slow."
  },
  {
   "id": "gb10-1u-gemma-4-31b-bf16",
   "model": "Gemma-4-31B-it",
   "model_slug": "gemma-4",
   "hardware": "1x GB10 (DGX Spark class)",
   "hardware_family": "gb10",
   "units": 1,
   "engine": "vLLM 0.29.1rc1.dev513",
   "quant": "BF16",
   "measured": "2026-09-23",
   "metrics": {
    "decode_1u_prose": {
     "value": 3.8,
     "source": "fleet-bench/2026-09-23-gb10-tp1-spark1-single-unit-survey.json",
     "quote": "\"rank\": 4, \"model\": \"gemma-4-31B-it bf16 (vLLM)\", \"family\": \"dense 31B\", \"single_stream_tok_s\": 3.8, \"peak_aggregate_tok_s\": 101.4, \"at_concurrency\": 32,",
     "json_key": "single_stream_tok_s"
    },
    "peak_agg_decode": {
     "value": 101.4,
     "source": "fleet-bench/2026-09-23-gb10-tp1-spark1-single-unit-survey.json",
     "quote": "\"rank\": 4, \"model\": \"gemma-4-31B-it bf16 (vLLM)\", \"family\": \"dense 31B\", \"single_stream_tok_s\": 3.8, \"peak_aggregate_tok_s\": 101.4, \"at_concurrency\": 32,",
     "json_key": "peak_aggregate_tok_s"
    }
   },
   "notes": "Fits one unit (63.2 GiB consumed) and batches cleanly, but 3.8 tok/s per stream is unusable interactively; bf16 on one GB10 is a waste of the box."
  },
  {
   "id": "gb10-1u-gpt-oss-120b-mxfp4",
   "model": "gpt-oss-120b",
   "model_slug": "gpt-oss-120b",
   "hardware": "1x GB10 (DGX Spark class)",
   "hardware_family": "gb10",
   "units": 1,
   "engine": "vLLM nightly",
   "quant": "MXFP4",
   "measured": "2026-10-02",
   "metrics": {
    "decode_1u_prose": {
     "value": 35.1,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| gpt-oss-120b MXFP4 (5.1B active) | vLLM nightly, no spec | 8/8 | 35.1 / 35.0 | 290.5 | 16 (198.0) | 4,900 / 4,037 / n/a (131k ctx) | 114 / 121 | 1.63 | 7.5 / 42 |",
     "column": "1 user prose / code",
     "part": 0
    },
    "decode_1u_code": {
     "value": 35.0,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| gpt-oss-120b MXFP4 (5.1B active) | vLLM nightly, no spec | 8/8 | 35.1 / 35.0 | 290.5 | 16 (198.0) | 4,900 / 4,037 / n/a (131k ctx) | 114 / 121 | 1.63 | 7.5 / 42 |",
     "column": "1 user prose / code",
     "part": 1
    },
    "agg_decode_32u": {
     "value": 290.5,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| gpt-oss-120b MXFP4 (5.1B active) | vLLM nightly, no spec | 8/8 | 35.1 / 35.0 | 290.5 | 16 (198.0) | 4,900 / 4,037 / n/a (131k ctx) | 114 / 121 | 1.63 | 7.5 / 42 |",
     "column": "32 users"
    },
    "ceiling_users": {
     "value": 16,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| gpt-oss-120b MXFP4 (5.1B active) | vLLM nightly, no spec | 8/8 | 35.1 / 35.0 | 290.5 | 16 (198.0) | 4,900 / 4,037 / n/a (131k ctx) | 114 / 121 | 1.63 | 7.5 / 42 |",
     "column": "ceiling users (agg)",
     "part": 0
    },
    "agg_decode_at_ceiling": {
     "value": 198.0,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| gpt-oss-120b MXFP4 (5.1B active) | vLLM nightly, no spec | 8/8 | 35.1 / 35.0 | 290.5 | 16 (198.0) | 4,900 / 4,037 / n/a (131k ctx) | 114 / 121 | 1.63 | 7.5 / 42 |",
     "column": "ceiling users (agg)",
     "part": 1
    },
    "prefill_8k": {
     "value": 4900,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| gpt-oss-120b MXFP4 (5.1B active) | vLLM nightly, no spec | 8/8 | 35.1 / 35.0 | 290.5 | 16 (198.0) | 4,900 / 4,037 / n/a (131k ctx) | 114 / 121 | 1.63 | 7.5 / 42 |",
     "column": "prefill 8k / 32k / 128k",
     "part": 0
    },
    "prefill_32k": {
     "value": 4037,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| gpt-oss-120b MXFP4 (5.1B active) | vLLM nightly, no spec | 8/8 | 35.1 / 35.0 | 290.5 | 16 (198.0) | 4,900 / 4,037 / n/a (131k ctx) | 114 / 121 | 1.63 | 7.5 / 42 |",
     "column": "prefill 8k / 32k / 128k",
     "part": 1
    },
    "tok_per_joule": {
     "value": 1.63,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| gpt-oss-120b MXFP4 (5.1B active) | vLLM nightly, no spec | 8/8 | 35.1 / 35.0 | 290.5 | 16 (198.0) | 4,900 / 4,037 / n/a (131k ctx) | 114 / 121 | 1.63 | 7.5 / 42 |",
     "column": "tok/J at ceiling"
    },
    "context_verified": {
     "value": 120755,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| gpt-oss-120b | 1 | 131,072 | 131,072 | 1,757,899 | 13.4 | 120,755: pass, 73 s | 35.7 / 26.8 at 117,364 | 51 s | 16 |",
     "column": "needle at 90% (prompt tokens, result, request time)",
     "part": 0
    },
    "cmmc_rag_strict_pct": {
     "value": 77,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| gpt-oss-120b (1) | 77 (s21; s22-23 cut by the outage) | 82 | 1 | 10/10 | 8 | 240/244 | 40.0 | 675 | 80 (0 fab) | 78 | 37 |",
     "column": "agentic strict /93 (seeds)",
     "part": 0
    }
   },
   "notes": "Slowest MoE tier-1 model; served context is capped at the 131,072-token model limit. Tool hygiene is clean (0% hallucinated tool calls)."
  },
  {
   "id": "gb10-1u-nemotron-3-5-lightning-nvfp4",
   "model": "Nemotron-3.5-Lightning-30B-A3B",
   "model_slug": "nemotron-3-5-lightning",
   "hardware": "1x GB10 (DGX Spark class)",
   "hardware_family": "gb10",
   "units": 1,
   "engine": "vLLM nightly",
   "quant": "NVFP4",
   "measured": "2026-10-02",
   "metrics": {
    "decode_1u_prose": {
     "value": 90.2,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| Nemotron-3.5-Lightning-30B-A3B NVFP4 | vLLM nightly, DSpark k3 (vendor DGX Spark recipe) | 7/8 + think-on recheck | 90.2 / 114.2 | 608.0 | 64 (795.4) | 6,859 / 6,431 / 4,632 | 111 / 126 | 6.30 | 19.2 / 169 |",
     "column": "1 user prose / code",
     "part": 0
    },
    "decode_1u_code": {
     "value": 114.2,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| Nemotron-3.5-Lightning-30B-A3B NVFP4 | vLLM nightly, DSpark k3 (vendor DGX Spark recipe) | 7/8 + think-on recheck | 90.2 / 114.2 | 608.0 | 64 (795.4) | 6,859 / 6,431 / 4,632 | 111 / 126 | 6.30 | 19.2 / 169 |",
     "column": "1 user prose / code",
     "part": 1
    },
    "agg_decode_32u": {
     "value": 608.0,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| Nemotron-3.5-Lightning-30B-A3B NVFP4 | vLLM nightly, DSpark k3 (vendor DGX Spark recipe) | 7/8 + think-on recheck | 90.2 / 114.2 | 608.0 | 64 (795.4) | 6,859 / 6,431 / 4,632 | 111 / 126 | 6.30 | 19.2 / 169 |",
     "column": "32 users"
    },
    "ceiling_users": {
     "value": 64,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| Nemotron-3.5-Lightning-30B-A3B NVFP4 | vLLM nightly, DSpark k3 (vendor DGX Spark recipe) | 7/8 + think-on recheck | 90.2 / 114.2 | 608.0 | 64 (795.4) | 6,859 / 6,431 / 4,632 | 111 / 126 | 6.30 | 19.2 / 169 |",
     "column": "ceiling users (agg)",
     "part": 0
    },
    "agg_decode_at_ceiling": {
     "value": 795.4,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| Nemotron-3.5-Lightning-30B-A3B NVFP4 | vLLM nightly, DSpark k3 (vendor DGX Spark recipe) | 7/8 + think-on recheck | 90.2 / 114.2 | 608.0 | 64 (795.4) | 6,859 / 6,431 / 4,632 | 111 / 126 | 6.30 | 19.2 / 169 |",
     "column": "ceiling users (agg)",
     "part": 1
    },
    "prefill_8k": {
     "value": 6859,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| Nemotron-3.5-Lightning-30B-A3B NVFP4 | vLLM nightly, DSpark k3 (vendor DGX Spark recipe) | 7/8 + think-on recheck | 90.2 / 114.2 | 608.0 | 64 (795.4) | 6,859 / 6,431 / 4,632 | 111 / 126 | 6.30 | 19.2 / 169 |",
     "column": "prefill 8k / 32k / 128k",
     "part": 0
    },
    "prefill_32k": {
     "value": 6431,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| Nemotron-3.5-Lightning-30B-A3B NVFP4 | vLLM nightly, DSpark k3 (vendor DGX Spark recipe) | 7/8 + think-on recheck | 90.2 / 114.2 | 608.0 | 64 (795.4) | 6,859 / 6,431 / 4,632 | 111 / 126 | 6.30 | 19.2 / 169 |",
     "column": "prefill 8k / 32k / 128k",
     "part": 1
    },
    "prefill_128k": {
     "value": 4632,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| Nemotron-3.5-Lightning-30B-A3B NVFP4 | vLLM nightly, DSpark k3 (vendor DGX Spark recipe) | 7/8 + think-on recheck | 90.2 / 114.2 | 608.0 | 64 (795.4) | 6,859 / 6,431 / 4,632 | 111 / 126 | 6.30 | 19.2 / 169 |",
     "column": "prefill 8k / 32k / 128k",
     "part": 2
    },
    "tok_per_joule": {
     "value": 6.3,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| Nemotron-3.5-Lightning-30B-A3B NVFP4 | vLLM nightly, DSpark k3 (vendor DGX Spark recipe) | 7/8 + think-on recheck | 90.2 / 114.2 | 608.0 | 64 (795.4) | 6,859 / 6,431 / 4,632 | 111 / 126 | 6.30 | 19.2 / 169 |",
     "column": "tok/J at ceiling"
    }
   },
   "notes": "The 09-25 speed battery was withheld (120k needle gate failure) and the 09-23 survey number (72.1 tok/s) is superseded by this 10-02 clean re-measure. Single-user speed leader on one GB10; quality is the weakest of the tier-1 MoE models (tools 29/38, long context 10/26)."
  },
  {
   "id": "gb10-1u-qwen3-6-35b-a3b-nvfp4",
   "model": "Qwen3.6-35B-A3B",
   "model_slug": "qwen3-6-35b-a3b",
   "hardware": "1x GB10 (DGX Spark class)",
   "hardware_family": "gb10",
   "units": 1,
   "engine": "vLLM nightly",
   "quant": "NVFP4",
   "measured": "2026-10-02",
   "metrics": {
    "decode_1u_prose": {
     "value": 88.2,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| Qwen3.6-35B-A3B NVFP4 (nvidia) | vLLM nightly, RedHat DSpark k8 | 8/8 | 88.2 / 115.7 | 452.9 | 48 (507.6) | 7,492 / 6,021 / 3,117 | 116 / 140 | 3.62 | 18.8 / 108 |",
     "column": "1 user prose / code",
     "part": 0
    },
    "decode_1u_code": {
     "value": 115.7,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| Qwen3.6-35B-A3B NVFP4 (nvidia) | vLLM nightly, RedHat DSpark k8 | 8/8 | 88.2 / 115.7 | 452.9 | 48 (507.6) | 7,492 / 6,021 / 3,117 | 116 / 140 | 3.62 | 18.8 / 108 |",
     "column": "1 user prose / code",
     "part": 1
    },
    "agg_decode_32u": {
     "value": 452.9,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| Qwen3.6-35B-A3B NVFP4 (nvidia) | vLLM nightly, RedHat DSpark k8 | 8/8 | 88.2 / 115.7 | 452.9 | 48 (507.6) | 7,492 / 6,021 / 3,117 | 116 / 140 | 3.62 | 18.8 / 108 |",
     "column": "32 users"
    },
    "ceiling_users": {
     "value": 48,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| Qwen3.6-35B-A3B NVFP4 (nvidia) | vLLM nightly, RedHat DSpark k8 | 8/8 | 88.2 / 115.7 | 452.9 | 48 (507.6) | 7,492 / 6,021 / 3,117 | 116 / 140 | 3.62 | 18.8 / 108 |",
     "column": "ceiling users (agg)",
     "part": 0
    },
    "agg_decode_at_ceiling": {
     "value": 507.6,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| Qwen3.6-35B-A3B NVFP4 (nvidia) | vLLM nightly, RedHat DSpark k8 | 8/8 | 88.2 / 115.7 | 452.9 | 48 (507.6) | 7,492 / 6,021 / 3,117 | 116 / 140 | 3.62 | 18.8 / 108 |",
     "column": "ceiling users (agg)",
     "part": 1
    },
    "prefill_8k": {
     "value": 7492,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| Qwen3.6-35B-A3B NVFP4 (nvidia) | vLLM nightly, RedHat DSpark k8 | 8/8 | 88.2 / 115.7 | 452.9 | 48 (507.6) | 7,492 / 6,021 / 3,117 | 116 / 140 | 3.62 | 18.8 / 108 |",
     "column": "prefill 8k / 32k / 128k",
     "part": 0
    },
    "prefill_32k": {
     "value": 6021,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| Qwen3.6-35B-A3B NVFP4 (nvidia) | vLLM nightly, RedHat DSpark k8 | 8/8 | 88.2 / 115.7 | 452.9 | 48 (507.6) | 7,492 / 6,021 / 3,117 | 116 / 140 | 3.62 | 18.8 / 108 |",
     "column": "prefill 8k / 32k / 128k",
     "part": 1
    },
    "prefill_128k": {
     "value": 3117,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| Qwen3.6-35B-A3B NVFP4 (nvidia) | vLLM nightly, RedHat DSpark k8 | 8/8 | 88.2 / 115.7 | 452.9 | 48 (507.6) | 7,492 / 6,021 / 3,117 | 116 / 140 | 3.62 | 18.8 / 108 |",
     "column": "prefill 8k / 32k / 128k",
     "part": 2
    },
    "tok_per_joule": {
     "value": 3.62,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| Qwen3.6-35B-A3B NVFP4 (nvidia) | vLLM nightly, RedHat DSpark k8 | 8/8 | 88.2 / 115.7 | 452.9 | 48 (507.6) | 7,492 / 6,021 / 3,117 | 116 / 140 | 3.62 | 18.8 / 108 |",
     "column": "tok/J at ceiling"
    },
    "x_cmmc_rag_agentic_strict": {
     "value": 0.735,
     "source": "fleet-bench/2026-10-03-cmmc-rag-v3-results.md",
     "quote": "| GB10 tier-1 (GB10 agent): qwen3.6-35b-a3b / gemma-4-26b-a4b / gpt-oss-120b / nemotron-3.5 / qwen3.8-27b | | | | 0.735 / 0.631 / 0.796 / 0.645 / 0.946 | 0.935 / 0.892 / 0.860 / 0.871 / 0.903 | 0.882 / 0.849 / 0.839 / 0.828 / 0.892 | 0.387 / 0.355 / 0.398 / 0.280 / 0.204 | | 1 each |",
     "label": "cmmc-rag-v3 agentic strict, one GB10, seed 21",
     "column": "Agentic",
     "part": 0
    }
   },
   "notes": "Matches Nemotron's speed with better tools, but closed-book CMMC recall (1-6 of 35) and long-document answers that exhaust the 6,144-token output budget while still reasoning rule it out for compliance or long-document work as served."
  },
  {
   "id": "gb10-1u-qwen3-6-35b-a3b-q4km-ollama",
   "model": "Qwen3.6-35B-A3B",
   "model_slug": "qwen3-6-35b-a3b",
   "hardware": "1x GB10 (DGX Spark class)",
   "hardware_family": "gb10",
   "units": 1,
   "engine": "Ollama 0.21.2",
   "quant": "Q4_K_M (GGUF)",
   "measured": "2026-09-23",
   "metrics": {
    "decode_1u_prose": {
     "value": 42.2,
     "source": "fleet-bench/2026-09-23-gb10-tp1-spark1-single-unit-survey.json",
     "quote": "\"rank\": 3, \"model\": \"qwen3.6:35b A3B Q4_K_M (ollama)\", \"family\": \"MoE 36B/3B-active\", \"single_stream_tok_s\": 42.2, \"peak_aggregate_tok_s\": 57.9, \"at_concurrency\": 8,",
     "json_key": "single_stream_tok_s"
    },
    "peak_agg_decode": {
     "value": 57.9,
     "source": "fleet-bench/2026-09-23-gb10-tp1-spark1-single-unit-survey.json",
     "quote": "\"rank\": 3, \"model\": \"qwen3.6:35b A3B Q4_K_M (ollama)\", \"family\": \"MoE 36B/3B-active\", \"single_stream_tok_s\": 42.2, \"peak_aggregate_tok_s\": 57.9, \"at_concurrency\": 8,",
     "json_key": "peak_aggregate_tok_s"
    }
   },
   "notes": "Fast for one user but Ollama refuses to batch this architecture, so it is a single-seat model only; do not put it behind a queue."
  },
  {
   "id": "gb10-1u-qwen3-8-27b-nvfp4",
   "model": "Qwen3.8-27B",
   "model_slug": "qwen3-8-27b",
   "hardware": "1x GB10 (DGX Spark class)",
   "hardware_family": "gb10",
   "units": 1,
   "engine": "vLLM nightly",
   "quant": "NVFP4",
   "measured": "2026-10-02",
   "metrics": {
    "decode_1u_prose": {
     "value": 24.4,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| Qwen3.8-27B NVFP4 (DENSE reference) | vLLM nightly, MTP3 (09-25 args) | 7/8 + recheck | 24.4 / 32.3 | 316.6 | 32 (316.6) | see JSON | 110 / 132 | 2.39 | 5.2 / 67 |",
     "column": "1 user prose / code",
     "part": 0
    },
    "decode_1u_code": {
     "value": 32.3,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| Qwen3.8-27B NVFP4 (DENSE reference) | vLLM nightly, MTP3 (09-25 args) | 7/8 + recheck | 24.4 / 32.3 | 316.6 | 32 (316.6) | see JSON | 110 / 132 | 2.39 | 5.2 / 67 |",
     "column": "1 user prose / code",
     "part": 1
    },
    "agg_decode_32u": {
     "value": 316.6,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| Qwen3.8-27B NVFP4 (DENSE reference) | vLLM nightly, MTP3 (09-25 args) | 7/8 + recheck | 24.4 / 32.3 | 316.6 | 32 (316.6) | see JSON | 110 / 132 | 2.39 | 5.2 / 67 |",
     "column": "32 users"
    },
    "ceiling_users": {
     "value": 32,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| Qwen3.8-27B NVFP4 (DENSE reference) | vLLM nightly, MTP3 (09-25 args) | 7/8 + recheck | 24.4 / 32.3 | 316.6 | 32 (316.6) | see JSON | 110 / 132 | 2.39 | 5.2 / 67 |",
     "column": "ceiling users (agg)",
     "part": 0
    },
    "tok_per_joule": {
     "value": 2.39,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| Qwen3.8-27B NVFP4 (DENSE reference) | vLLM nightly, MTP3 (09-25 args) | 7/8 + recheck | 24.4 / 32.3 | 316.6 | 32 (316.6) | see JSON | 110 / 132 | 2.39 | 5.2 / 67 |",
     "column": "tok/J at ceiling"
    },
    "context_verified": {
     "value": 243392,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| Qwen3.8-27B (dense) | 1 | 262,144 | 262,144 | 992,020 | 3.8 | 243,392: pass, 335 s | 24.9 / 8.2 at 235,329 | 299 s | 32 |",
     "column": "needle at 90% (prompt tokens, result, request time)",
     "part": 0
    }
   },
   "notes": "Dense reference. Reproduces the 09-25 result (24.4 vs 23.6 prose) and remains the best tool/agent/long-context model on one unit, at a quarter of Nemotron's speed. Closed-book CMMC recall is effectively absent (1-4 of 35)."
  },
  {
   "id": "gb10-8u-mimo-v2-6-pro-rl-sglang",
   "model": "MiMo-V2.6-Pro-RL",
   "model_slug": "mimo-v2-6-pro-rl",
   "hardware": "8x GB10, SGLang rhys101 a14 with DFlash-8 adaptive",
   "hardware_family": "gb10",
   "units": 8,
   "engine": "SGLang",
   "quant": "published kit weights",
   "measured": "2026-10-03",
   "metrics": {
    "decode_1u_prose": {
     "value": 41.3,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| Our battery prose 1 / 2 / 4 / 8 users (agg) | 41.3 / 55.3 / 65.8 / 93.8 | 40.1 / 53.5 / 62.5 / 89.1 | - |",
     "column": "MSIs unlocked (Sparks 2300)",
     "part": 0
    },
    "decode_1u_code": {
     "value": 54.5,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| Our battery code 1 / 2 / 4 / 8 users | 54.5 / 69.4 / 75.4 / 127.5 | 58.0 / 68.8 / 76.7 / 129.4 | - |",
     "column": "MSIs unlocked (Sparks 2300)",
     "part": 0
    },
    "agg_decode_8u": {
     "value": 93.8,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| Our battery prose 1 / 2 / 4 / 8 users (agg) | 41.3 / 55.3 / 65.8 / 93.8 | 40.1 / 53.5 / 62.5 / 89.1 | - |",
     "column": "MSIs unlocked (Sparks 2300)",
     "part": 3
    },
    "prefill_8k": {
     "value": 2009,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| Prefill 8k / 32k / 128k / 257k | 2,009 / 1,722 / 1,600 / 1,378 | 2,002 / 1,956 / 1,687 / 1,441 | - |",
     "column": "MSIs unlocked (Sparks 2300)",
     "part": 0
    },
    "prefill_32k": {
     "value": 1722,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| Prefill 8k / 32k / 128k / 257k | 2,009 / 1,722 / 1,600 / 1,378 | 2,002 / 1,956 / 1,687 / 1,441 | - |",
     "column": "MSIs unlocked (Sparks 2300)",
     "part": 1
    },
    "prefill_128k": {
     "value": 1600,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| Prefill 8k / 32k / 128k / 257k | 2,009 / 1,722 / 1,600 / 1,378 | 2,002 / 1,956 / 1,687 / 1,441 | - |",
     "column": "MSIs unlocked (Sparks 2300)",
     "part": 2
    },
    "power_w": {
     "value": 1169,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| Pod wall W (8 outlets) at 1 / 8 users, prose | 1,169 / 1,179 | 1,081 / 1,070 | - |",
     "column": "MSIs unlocked (Sparks 2300)",
     "part": 0
    },
    "tok_per_joule": {
     "value": 0.08,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| tok/J at 8 users (prose / code) | 0.080 / 0.108 | 0.083 / 0.115 | - |",
     "column": "MSIs unlocked (Sparks 2300)",
     "part": 0
    },
    "x_cmmc_closedbook": {
     "value": 24,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "Quality (vendor sampling, thinking on, as the kit specifies; MSIs-unlocked run): tools 38/38, agent 15/15, **CMMC 24 /35",
     "label": "closed-book CMMC score of 35, thinking on"
    }
   },
   "notes": "Low-concurrency profile (8 slots: 94 prose / 128 code tok/s aggregate at 8 users) drawing about 1.1-1.2 kW. Reproduces its published speed within 5% (1 stream) and 2% (8 streams); the only tier-8 model here with GLM-class closed-book CMMC."
  },
  {
   "id": "gb10-tp2-deepseek-v4-1-flash-1017-exl3-mia",
   "model": "DeepSeek-V4.1-Flash-1017",
   "model_slug": "deepseek-v4-1-flash",
   "hardware": "2x GB10, TP2 (spark1+spark2)",
   "hardware_family": "gb10",
   "units": 2,
   "engine": "TensorFold native Zig (MiaAI recipe)",
   "quant": "EAQ v1 EXL3 2.9 bpw",
   "measured": "2026-10-23",
   "metrics": {
    "decode_1u_prose": {
     "value": 54.9,
     "source": "fleet-bench/t1023-mia-ds1017/SUMMARY.md",
     "quote": "| Real-text C1 prose tok/s | 54.9 | 54.9 | n/a | 177 |",
     "column": "Stock"
    },
    "decode_1u_code": {
     "value": 78.9,
     "source": "fleet-bench/t1023-mia-ds1017/SUMMARY.md",
     "quote": "| Real-text C1 code tok/s | 78.9-89.1 | 78.9 | n/a | 292 |",
     "column": "Stock",
     "part": 0
    },
    "prefill_8k": {
     "value": 2045,
     "source": "fleet-bench/t1023-mia-ds1017/SUMMARY.md",
     "quote": "| Prefill 8K/32K/128K tok/s | 2045/2066/2043 | 2043/2171/2063 | 1755/.../1713 | - |",
     "column": "Stock",
     "part": 0
    },
    "prefill_32k": {
     "value": 2066,
     "source": "fleet-bench/t1023-mia-ds1017/SUMMARY.md",
     "quote": "| Prefill 8K/32K/128K tok/s | 2045/2066/2043 | 2043/2171/2063 | 1755/.../1713 | - |",
     "column": "Stock",
     "part": 1
    },
    "prefill_128k": {
     "value": 2043,
     "source": "fleet-bench/t1023-mia-ds1017/SUMMARY.md",
     "quote": "| Prefill 8K/32K/128K tok/s | 2045/2066/2043 | 2043/2171/2063 | 1755/.../1713 | - |",
     "column": "Stock",
     "part": 2
    },
    "toolcall_pass": {
     "value": 20,
     "source": "fleet-bench/t1023-mia-ds1017/SUMMARY.md",
     "quote": "| toolv2 (of 27) | 20 | 21 | 20 | - |",
     "column": "Stock"
    },
    "x_cmmc_rag_seed21_correct": {
     "value": 87,
     "source": "fleet-bench/t1023-mia-ds1017/SUMMARY.md",
     "quote": "| cmmc-rag-v3 seed21 acc/strict/fab | 87/84/4 | 87/84/4 | 91/90/3 | - |",
     "label": "cmmc-rag-v3 agentic seed 21, correct of 93",
     "column": "Stock",
     "part": 0
    }
   },
   "notes": "Stock Mia EAQ quant. The ablated build measured the same speed and quality within noise. Quant quality sits a touch below the t1018 TensorFold run on CMMC (one seed each); Mia labels the quant experimental."
  },
  {
   "id": "gb10-tp2-deepseek-v4-1-flash-1017-exl3-tensorfold",
   "model": "DeepSeek-V4.1-Flash-1017",
   "model_slug": "deepseek-v4-1-flash",
   "hardware": "2x GB10, TP2 (spark1+spark2)",
   "hardware_family": "gb10",
   "units": 2,
   "engine": "TensorFold (Tony's)",
   "quant": "EXL3 2.9 bpw",
   "measured": "2026-10-18",
   "metrics": {
    "decode_1u_prose": {
     "value": 65.7,
     "source": "fleet-bench/t1018-ds-vs-glm/SUMMARY.md",
     "quote": "| prose agg tok/s, 1 stream | **65.7** | 28.7 | 22.6 |",
     "column": "TensorFold DS-V4.1 2.9 bpw"
    },
    "prefill_8k": {
     "value": 1713,
     "source": "fleet-bench/t1018-ds-vs-glm/SUMMARY.md",
     "quote": "| cold prefill tok/s | 1,713-1,887 | 1,648-1,812 | - |",
     "column": "TensorFold DS-V4.1 2.9 bpw",
     "part": 0
    },
    "toolcall_pass": {
     "value": 20,
     "source": "fleet-bench/t1018-ds-vs-glm/SUMMARY.md",
     "quote": "| toolv2 (27) | 20 | 18 | no 2-Spark number (H200 NVFP4: 24) |",
     "column": "TensorFold DS-V4.1 2.9 bpw"
    },
    "x_cmmc_rag_seed21_correct": {
     "value": 91,
     "source": "fleet-bench/t1018-ds-vs-glm/SUMMARY.md",
     "quote": "| cmmc-rag-v3 agentic seed 21: correct / strict / fab | **91 / 90** / 3 | 75 / 70 / 0 | 91 / 88 / 1 |",
     "label": "cmmc-rag-v3 agentic seed 21, correct of 93",
     "column": "TensorFold DS-V4.1 2.9 bpw",
     "part": 0
    },
    "x_cmmc_rag_seed21_strict": {
     "value": 90,
     "source": "fleet-bench/t1018-ds-vs-glm/SUMMARY.md",
     "quote": "| cmmc-rag-v3 agentic seed 21: correct / strict / fab | **91 / 90** / 3 | 75 / 70 / 0 | 91 / 88 / 1 |",
     "label": "cmmc-rag-v3 agentic seed 21, strict of 93",
     "column": "TensorFold DS-V4.1 2.9 bpw",
     "part": 1
    },
    "x_agg_decode_4u": {
     "value": 125.6,
     "source": "fleet-bench/t1018-ds-vs-glm/SUMMARY.md",
     "quote": "| agg tok/s, 4 streams | **125.6** | 56.0 | - |",
     "label": "aggregate decode at 4 streams",
     "column": "TensorFold DS-V4.1 2.9 bpw"
    }
   }
  },
  {
   "id": "gb10-tp2-deepseek-v4-flash-vision-exp-fp8",
   "model": "DeepSeek-V4-Flash-Vision-Exp",
   "model_slug": "deepseek-v4-1-flash",
   "hardware": "2x GB10, TP2 (spark3+4)",
   "hardware_family": "gb10",
   "units": 2,
   "engine": "vLLM, Mia's Anemll DSpark kit",
   "quant": "FP8",
   "measured": "2026-10-02",
   "metrics": {
    "decode_1u_prose": {
     "value": 47.8,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| DeepSeek-V4-Flash-Vision-Exp | Mia's Anemll DSpark kit, 6 seqs, 1M ctx | spark3+4, 2300 | 8/8 | 47.8 / 43.5 | 6 (117.0; profile cap) | 117.0 at 6 | 1,732 / 1,682 / 1,484 | 269 / 288 | 0.41 |",
     "column": "1 user prose / code",
     "part": 0
    },
    "decode_1u_code": {
     "value": 43.5,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| DeepSeek-V4-Flash-Vision-Exp | Mia's Anemll DSpark kit, 6 seqs, 1M ctx | spark3+4, 2300 | 8/8 | 47.8 / 43.5 | 6 (117.0; profile cap) | 117.0 at 6 | 1,732 / 1,682 / 1,484 | 269 / 288 | 0.41 |",
     "column": "1 user prose / code",
     "part": 1
    },
    "ceiling_users": {
     "value": 6,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| DeepSeek-V4-Flash-Vision-Exp | Mia's Anemll DSpark kit, 6 seqs, 1M ctx | spark3+4, 2300 | 8/8 | 47.8 / 43.5 | 6 (117.0; profile cap) | 117.0 at 6 | 1,732 / 1,682 / 1,484 | 269 / 288 | 0.41 |",
     "column": "users at floor (agg)",
     "part": 0
    },
    "agg_decode_at_ceiling": {
     "value": 117.0,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| DeepSeek-V4-Flash-Vision-Exp | Mia's Anemll DSpark kit, 6 seqs, 1M ctx | spark3+4, 2300 | 8/8 | 47.8 / 43.5 | 6 (117.0; profile cap) | 117.0 at 6 | 1,732 / 1,682 / 1,484 | 269 / 288 | 0.41 |",
     "column": "top-of-ladder agg",
     "part": 0
    },
    "prefill_8k": {
     "value": 1732,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| DeepSeek-V4-Flash-Vision-Exp | Mia's Anemll DSpark kit, 6 seqs, 1M ctx | spark3+4, 2300 | 8/8 | 47.8 / 43.5 | 6 (117.0; profile cap) | 117.0 at 6 | 1,732 / 1,682 / 1,484 | 269 / 288 | 0.41 |",
     "column": "prefill 8k / 32k / 128k",
     "part": 0
    },
    "prefill_32k": {
     "value": 1682,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| DeepSeek-V4-Flash-Vision-Exp | Mia's Anemll DSpark kit, 6 seqs, 1M ctx | spark3+4, 2300 | 8/8 | 47.8 / 43.5 | 6 (117.0; profile cap) | 117.0 at 6 | 1,732 / 1,682 / 1,484 | 269 / 288 | 0.41 |",
     "column": "prefill 8k / 32k / 128k",
     "part": 1
    },
    "prefill_128k": {
     "value": 1484,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| DeepSeek-V4-Flash-Vision-Exp | Mia's Anemll DSpark kit, 6 seqs, 1M ctx | spark3+4, 2300 | 8/8 | 47.8 / 43.5 | 6 (117.0; profile cap) | 117.0 at 6 | 1,732 / 1,682 / 1,484 | 269 / 288 | 0.41 |",
     "column": "prefill 8k / 32k / 128k",
     "part": 2
    },
    "tok_per_joule": {
     "value": 0.41,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| DeepSeek-V4-Flash-Vision-Exp | Mia's Anemll DSpark kit, 6 seqs, 1M ctx | spark3+4, 2300 | 8/8 | 47.8 / 43.5 | 6 (117.0; profile cap) | 117.0 at 6 | 1,732 / 1,682 / 1,484 | 269 / 288 | 0.41 |",
     "column": "tok/J at ceiling"
    }
   },
   "notes": "Best two-unit single-user model (48 tok/s prose) and the best tier-2 quality short of GLM (tools 38, agent 15, long 26/26), but the kit is a 6-sequence profile capped at 6 users at the floor. Serves 1,048,576-token context; a 120k needle passed."
  },
  {
   "id": "gb10-tp2-glm-5-3-flash-nvfp4-dflash2",
   "model": "GLM-5.3-Flash",
   "model_slug": "glm-5-3-flash",
   "hardware": "2x GB10, TP2 (msi3+4)",
   "hardware_family": "gb10",
   "units": 2,
   "engine": "vLLM, tonyd2wild v11 DFlash2",
   "quant": "NVFP4",
   "measured": "2026-10-02",
   "metrics": {
    "decode_1u_prose": {
     "value": 24.1,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| GLM-5.3-Flash NVFP4 (nvidia) | tonyd2wild v11, DFlash2 k7, 6 seqs, 8 GiB KV | msi3+4, 2300 | 8/8 | 24.1 / 26.8 | 6 (68.3; cap) | 68.3 | 1,604 / 987 / 1,171 | 265 / 279 | 0.23 |",
     "column": "1 user prose / code",
     "part": 0
    },
    "decode_1u_code": {
     "value": 26.8,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| GLM-5.3-Flash NVFP4 (nvidia) | tonyd2wild v11, DFlash2 k7, 6 seqs, 8 GiB KV | msi3+4, 2300 | 8/8 | 24.1 / 26.8 | 6 (68.3; cap) | 68.3 | 1,604 / 987 / 1,171 | 265 / 279 | 0.23 |",
     "column": "1 user prose / code",
     "part": 1
    },
    "ceiling_users": {
     "value": 6,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| GLM-5.3-Flash NVFP4 (nvidia) | tonyd2wild v11, DFlash2 k7, 6 seqs, 8 GiB KV | msi3+4, 2300 | 8/8 | 24.1 / 26.8 | 6 (68.3; cap) | 68.3 | 1,604 / 987 / 1,171 | 265 / 279 | 0.23 |",
     "column": "users at floor (agg)",
     "part": 0
    },
    "agg_decode_at_ceiling": {
     "value": 68.3,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| GLM-5.3-Flash NVFP4 (nvidia) | tonyd2wild v11, DFlash2 k7, 6 seqs, 8 GiB KV | msi3+4, 2300 | 8/8 | 24.1 / 26.8 | 6 (68.3; cap) | 68.3 | 1,604 / 987 / 1,171 | 265 / 279 | 0.23 |",
     "column": "top-of-ladder agg"
    },
    "prefill_8k": {
     "value": 1604,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| GLM-5.3-Flash NVFP4 (nvidia) | tonyd2wild v11, DFlash2 k7, 6 seqs, 8 GiB KV | msi3+4, 2300 | 8/8 | 24.1 / 26.8 | 6 (68.3; cap) | 68.3 | 1,604 / 987 / 1,171 | 265 / 279 | 0.23 |",
     "column": "prefill 8k / 32k / 128k",
     "part": 0
    },
    "prefill_32k": {
     "value": 987,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| GLM-5.3-Flash NVFP4 (nvidia) | tonyd2wild v11, DFlash2 k7, 6 seqs, 8 GiB KV | msi3+4, 2300 | 8/8 | 24.1 / 26.8 | 6 (68.3; cap) | 68.3 | 1,604 / 987 / 1,171 | 265 / 279 | 0.23 |",
     "column": "prefill 8k / 32k / 128k",
     "part": 1
    },
    "prefill_128k": {
     "value": 1171,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| GLM-5.3-Flash NVFP4 (nvidia) | tonyd2wild v11, DFlash2 k7, 6 seqs, 8 GiB KV | msi3+4, 2300 | 8/8 | 24.1 / 26.8 | 6 (68.3; cap) | 68.3 | 1,604 / 987 / 1,171 | 265 / 279 | 0.23 |",
     "column": "prefill 8k / 32k / 128k",
     "part": 2
    },
    "tok_per_joule": {
     "value": 0.23,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| GLM-5.3-Flash NVFP4 (nvidia) | tonyd2wild v11, DFlash2 k7, 6 seqs, 8 GiB KV | msi3+4, 2300 | 8/8 | 24.1 / 26.8 | 6 (68.3; cap) | 68.3 | 1,604 / 987 / 1,171 | 265 / 279 | 0.23 |",
     "column": "tok/J at ceiling"
    }
   },
   "notes": "Best quality on the pod (CMMC 26 of 35 closed-book, 0% hallucinated tool calls) but at TP2 it fits only with 6 sequences and 24-28 tok/s; long-context GLM belongs on 4 or more units. The 32k prefill value of 987 is flagged in the source as an outlier to treat as noise."
  },
  {
   "id": "gb10-tp2-mimo-v2-6-flash-rl-fp8",
   "model": "MiMo-V2.6-Flash-RL",
   "model_slug": "mimo-v2-6-flash",
   "hardware": "2x GB10, TP2 (spark3+4)",
   "hardware_family": "gb10",
   "units": 2,
   "engine": "vLLM",
   "quant": "FP8 (w8a8)",
   "measured": "2026-09-23",
   "metrics": {
    "decode_1u_prose": {
     "value": 26.6,
     "source": "fleet-bench/2026-09-23-gb10-tp2-spark34-mimo-v26-flash.json",
     "quote": "\"concurrency\": 1, \"ok\": 1, \"failed\": 0, \"wall_s\": 9.64, \"aggregate_tok_s\": 26.6, \"prompt_tokens_median\": 36,",
     "json_key": "aggregate_tok_s"
    },
    "peak_agg_decode": {
     "value": 205.4,
     "source": "fleet-bench/2026-09-23-gb10-tp2-spark34-mimo-v26-flash.json",
     "quote": "\"concurrency\": 32, \"ok\": 32, \"failed\": 0, \"wall_s\": 39.88, \"aggregate_tok_s\": 205.4,",
     "json_key": "aggregate_tok_s"
    }
   },
   "notes": "Speculative decoding off (non-DFlash pair baseline). Aggregate peaks at 205.4 tok/s at 32 users and falls to 177.0 at 48 as TTFT climbs past 40 s."
  },
  {
   "id": "gb10-tp2-qwen3-8-flash-next-nvfp4",
   "model": "Qwen3.8-Flash-Next",
   "model_slug": "qwen3-8-flash-next",
   "hardware": "2x GB10, TP2 (spark1+2)",
   "hardware_family": "gb10",
   "units": 2,
   "engine": "vLLM nightly, oxbyte MTP4, EP",
   "quant": "NVFP4",
   "measured": "2026-10-02",
   "metrics": {
    "decode_1u_prose": {
     "value": 35.2,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| Qwen3.8-Flash-Next NVFP4 | oxbyte, MTP4, EP, 64 seqs (10-01 serving config) | spark1+2, 2300 | 8/8 | 35.2 / 52.6 | 32 (329.6) | 329.6 at 32 | 3,089 / 3,069 / n/a (131k ctx) | 251 / 285 | 1.16 |",
     "column": "1 user prose / code",
     "part": 0
    },
    "decode_1u_code": {
     "value": 52.6,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| Qwen3.8-Flash-Next NVFP4 | oxbyte, MTP4, EP, 64 seqs (10-01 serving config) | spark1+2, 2300 | 8/8 | 35.2 / 52.6 | 32 (329.6) | 329.6 at 32 | 3,089 / 3,069 / n/a (131k ctx) | 251 / 285 | 1.16 |",
     "column": "1 user prose / code",
     "part": 1
    },
    "agg_decode_32u": {
     "value": 329.6,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| Qwen3.8-Flash-Next NVFP4 | oxbyte, MTP4, EP, 64 seqs (10-01 serving config) | spark1+2, 2300 | 8/8 | 35.2 / 52.6 | 32 (329.6) | 329.6 at 32 | 3,089 / 3,069 / n/a (131k ctx) | 251 / 285 | 1.16 |",
     "column": "users at floor (agg)",
     "part": 1
    },
    "ceiling_users": {
     "value": 32,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| Qwen3.8-Flash-Next NVFP4 | oxbyte, MTP4, EP, 64 seqs (10-01 serving config) | spark1+2, 2300 | 8/8 | 35.2 / 52.6 | 32 (329.6) | 329.6 at 32 | 3,089 / 3,069 / n/a (131k ctx) | 251 / 285 | 1.16 |",
     "column": "users at floor (agg)",
     "part": 0
    },
    "agg_decode_at_ceiling": {
     "value": 329.6,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| Qwen3.8-Flash-Next NVFP4 | oxbyte, MTP4, EP, 64 seqs (10-01 serving config) | spark1+2, 2300 | 8/8 | 35.2 / 52.6 | 32 (329.6) | 329.6 at 32 | 3,089 / 3,069 / n/a (131k ctx) | 251 / 285 | 1.16 |",
     "column": "top-of-ladder agg",
     "part": 0
    },
    "prefill_8k": {
     "value": 3089,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| Qwen3.8-Flash-Next NVFP4 | oxbyte, MTP4, EP, 64 seqs (10-01 serving config) | spark1+2, 2300 | 8/8 | 35.2 / 52.6 | 32 (329.6) | 329.6 at 32 | 3,089 / 3,069 / n/a (131k ctx) | 251 / 285 | 1.16 |",
     "column": "prefill 8k / 32k / 128k",
     "part": 0
    },
    "prefill_32k": {
     "value": 3069,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| Qwen3.8-Flash-Next NVFP4 | oxbyte, MTP4, EP, 64 seqs (10-01 serving config) | spark1+2, 2300 | 8/8 | 35.2 / 52.6 | 32 (329.6) | 329.6 at 32 | 3,089 / 3,069 / n/a (131k ctx) | 251 / 285 | 1.16 |",
     "column": "prefill 8k / 32k / 128k",
     "part": 1
    },
    "tok_per_joule": {
     "value": 1.16,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| Qwen3.8-Flash-Next NVFP4 | oxbyte, MTP4, EP, 64 seqs (10-01 serving config) | spark1+2, 2300 | 8/8 | 35.2 / 52.6 | 32 (329.6) | 329.6 at 32 | 3,089 / 3,069 / n/a (131k ctx) | 251 / 285 | 1.16 |",
     "column": "tok/J at ceiling"
    },
    "x_cmmc_rag_agentic_strict": {
     "value": 0.978,
     "source": "fleet-bench/2026-10-03-cmmc-rag-v3-results.md",
     "quote": "| **gb10-qwenfn-tp2** (Qwen3.8-Flash-Next NVFP4 @262k) | **0.978** (3) | 0.935 (1) | 0.892 (1) | 0.247 |",
     "label": "cmmc-rag-v3 agentic strict, TP2 pair, 3 seeds",
     "column": "Agentic",
     "part": 0
    }
   },
   "notes": "The two-unit serving model: 32 users at the floor at 330 tok/s aggregate. Its cmmc-rag-v3 agentic strict score of 0.978 ties the FP8 Qwen3.8-27B on one RTX PRO 6000 Blackwell GPU (0.971), making the pair the pod's quality reference as well as its throughput leader."
  },
  {
   "id": "gb10-tp4-deepseek-v4-1-flash-nvfp4-dspark",
   "model": "DeepSeek-V4-Flash-0731",
   "model_slug": "deepseek-v4-1-flash",
   "hardware": "4x GB10, TP4 (spark1-4)",
   "hardware_family": "gb10",
   "units": 4,
   "engine": "vLLM, AlexLJC/MiaAI kit, DSpark k5",
   "quant": "NVFP4",
   "measured": "2026-10-02",
   "metrics": {
    "decode_1u_prose": {
     "value": 54.1,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| DeepSeek-V4-Flash-0731, forum #1 kit at TP4, DSpark k5, 12 seqs | spark1-4, 2300 | 7/8 + think-on recheck | 54.1 / 94.2 | 120.9 / 173.6 / 202.0 at 12 (code 354.2 at 12) | 12 (202.0; profile cap) | 2,220 / 2,156 / 1,826 | 513 / 572 at 12 | 0.35 |",
     "column": "1 user prose / code",
     "part": 0
    },
    "decode_1u_code": {
     "value": 94.2,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| DeepSeek-V4-Flash-0731, forum #1 kit at TP4, DSpark k5, 12 seqs | spark1-4, 2300 | 7/8 + think-on recheck | 54.1 / 94.2 | 120.9 / 173.6 / 202.0 at 12 (code 354.2 at 12) | 12 (202.0; profile cap) | 2,220 / 2,156 / 1,826 | 513 / 572 at 12 | 0.35 |",
     "column": "1 user prose / code",
     "part": 1
    },
    "ceiling_users": {
     "value": 12,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| DeepSeek-V4-Flash-0731, forum #1 kit at TP4, DSpark k5, 12 seqs | spark1-4, 2300 | 7/8 + think-on recheck | 54.1 / 94.2 | 120.9 / 173.6 / 202.0 at 12 (code 354.2 at 12) | 12 (202.0; profile cap) | 2,220 / 2,156 / 1,826 | 513 / 572 at 12 | 0.35 |",
     "column": "users at floor (agg)",
     "part": 0
    },
    "agg_decode_at_ceiling": {
     "value": 202.0,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| DeepSeek-V4-Flash-0731, forum #1 kit at TP4, DSpark k5, 12 seqs | spark1-4, 2300 | 7/8 + think-on recheck | 54.1 / 94.2 | 120.9 / 173.6 / 202.0 at 12 (code 354.2 at 12) | 12 (202.0; profile cap) | 2,220 / 2,156 / 1,826 | 513 / 572 at 12 | 0.35 |",
     "column": "users at floor (agg)",
     "part": 1
    },
    "prefill_8k": {
     "value": 2220,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| DeepSeek-V4-Flash-0731, forum #1 kit at TP4, DSpark k5, 12 seqs | spark1-4, 2300 | 7/8 + think-on recheck | 54.1 / 94.2 | 120.9 / 173.6 / 202.0 at 12 (code 354.2 at 12) | 12 (202.0; profile cap) | 2,220 / 2,156 / 1,826 | 513 / 572 at 12 | 0.35 |",
     "column": "prefill 8k / 32k / 128k",
     "part": 0
    },
    "prefill_32k": {
     "value": 2156,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| DeepSeek-V4-Flash-0731, forum #1 kit at TP4, DSpark k5, 12 seqs | spark1-4, 2300 | 7/8 + think-on recheck | 54.1 / 94.2 | 120.9 / 173.6 / 202.0 at 12 (code 354.2 at 12) | 12 (202.0; profile cap) | 2,220 / 2,156 / 1,826 | 513 / 572 at 12 | 0.35 |",
     "column": "prefill 8k / 32k / 128k",
     "part": 1
    },
    "prefill_128k": {
     "value": 1826,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| DeepSeek-V4-Flash-0731, forum #1 kit at TP4, DSpark k5, 12 seqs | spark1-4, 2300 | 7/8 + think-on recheck | 54.1 / 94.2 | 120.9 / 173.6 / 202.0 at 12 (code 354.2 at 12) | 12 (202.0; profile cap) | 2,220 / 2,156 / 1,826 | 513 / 572 at 12 | 0.35 |",
     "column": "prefill 8k / 32k / 128k",
     "part": 2
    },
    "tok_per_joule": {
     "value": 0.35,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| DeepSeek-V4-Flash-0731, forum #1 kit at TP4, DSpark k5, 12 seqs | spark1-4, 2300 | 7/8 + think-on recheck | 54.1 / 94.2 | 120.9 / 173.6 / 202.0 at 12 (code 354.2 at 12) | 12 (202.0; profile cap) | 2,220 / 2,156 / 1,826 | 513 / 572 at 12 | 0.35 |",
     "column": "tok/J 32 users"
    }
   },
   "notes": "12-sequence profile capped at 12 users at the floor; code aggregate reaches 354.2 tok/s at 12 users. Serves 1,048,576-token context with a 7.65M-token KV pool."
  },
  {
   "id": "gb10-tp4-glm-5-3-flash-nvfp4-dflash2",
   "model": "GLM-5.3-Flash",
   "model_slug": "glm-5-3-flash",
   "hardware": "4x GB10, TP4 (msi1-4)",
   "hardware_family": "gb10",
   "units": 4,
   "engine": "vLLM, tonyd2wild v11 DFlash2 k7",
   "quant": "NVFP4",
   "measured": "2026-10-02",
   "metrics": {
    "decode_1u_prose": {
     "value": 44.9,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| GLM-5.3-Flash DFlash2 k7, 64 seqs, 1M ctx | msi1-4, 2300 | 8/8 | 44.9 / 49.9 | 96.1 / 135.9 / 174.3 / 258.6 | 16 (174.3) | 2,120 / 2,049 / 2,048 | 507 / 579 | 0.45 |",
     "column": "1 user prose / code",
     "part": 0
    },
    "decode_1u_code": {
     "value": 49.9,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| GLM-5.3-Flash DFlash2 k7, 64 seqs, 1M ctx | msi1-4, 2300 | 8/8 | 44.9 / 49.9 | 96.1 / 135.9 / 174.3 / 258.6 | 16 (174.3) | 2,120 / 2,049 / 2,048 | 507 / 579 | 0.45 |",
     "column": "1 user prose / code",
     "part": 1
    },
    "agg_decode_32u": {
     "value": 258.6,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| GLM-5.3-Flash DFlash2 k7, 64 seqs, 1M ctx | msi1-4, 2300 | 8/8 | 44.9 / 49.9 | 96.1 / 135.9 / 174.3 / 258.6 | 16 (174.3) | 2,120 / 2,049 / 2,048 | 507 / 579 | 0.45 |",
     "column": "4 / 8 / 16 / 32 users agg (prose)",
     "part": 3
    },
    "ceiling_users": {
     "value": 16,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| GLM-5.3-Flash DFlash2 k7, 64 seqs, 1M ctx | msi1-4, 2300 | 8/8 | 44.9 / 49.9 | 96.1 / 135.9 / 174.3 / 258.6 | 16 (174.3) | 2,120 / 2,049 / 2,048 | 507 / 579 | 0.45 |",
     "column": "users at floor (agg)",
     "part": 0
    },
    "agg_decode_at_ceiling": {
     "value": 174.3,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| GLM-5.3-Flash DFlash2 k7, 64 seqs, 1M ctx | msi1-4, 2300 | 8/8 | 44.9 / 49.9 | 96.1 / 135.9 / 174.3 / 258.6 | 16 (174.3) | 2,120 / 2,049 / 2,048 | 507 / 579 | 0.45 |",
     "column": "users at floor (agg)",
     "part": 1
    },
    "prefill_8k": {
     "value": 2120,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| GLM-5.3-Flash DFlash2 k7, 64 seqs, 1M ctx | msi1-4, 2300 | 8/8 | 44.9 / 49.9 | 96.1 / 135.9 / 174.3 / 258.6 | 16 (174.3) | 2,120 / 2,049 / 2,048 | 507 / 579 | 0.45 |",
     "column": "prefill 8k / 32k / 128k",
     "part": 0
    },
    "prefill_32k": {
     "value": 2049,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| GLM-5.3-Flash DFlash2 k7, 64 seqs, 1M ctx | msi1-4, 2300 | 8/8 | 44.9 / 49.9 | 96.1 / 135.9 / 174.3 / 258.6 | 16 (174.3) | 2,120 / 2,049 / 2,048 | 507 / 579 | 0.45 |",
     "column": "prefill 8k / 32k / 128k",
     "part": 1
    },
    "prefill_128k": {
     "value": 2048,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| GLM-5.3-Flash DFlash2 k7, 64 seqs, 1M ctx | msi1-4, 2300 | 8/8 | 44.9 / 49.9 | 96.1 / 135.9 / 174.3 / 258.6 | 16 (174.3) | 2,120 / 2,049 / 2,048 | 507 / 579 | 0.45 |",
     "column": "prefill 8k / 32k / 128k",
     "part": 2
    },
    "tok_per_joule": {
     "value": 0.45,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| GLM-5.3-Flash DFlash2 k7, 64 seqs, 1M ctx | msi1-4, 2300 | 8/8 | 44.9 / 49.9 | 96.1 / 135.9 / 174.3 / 258.6 | 16 (174.3) | 2,120 / 2,049 / 2,048 | 507 / 579 | 0.45 |",
     "column": "tok/J 32 users"
    }
   },
   "notes": "Serves 1,048,576-token context with a 3.9M-token KV pool and 3.7 full-length requests at once. Quality carried over from the 09-26 TP4 battery: CMMC 26.0 of 35, 0% hallucinated tool calls, long context 26/26."
  },
  {
   "id": "gb10-tp4-qwen3-8-flash-next-nvfp4",
   "model": "Qwen3.8-Flash-Next",
   "model_slug": "qwen3-8-flash-next",
   "hardware": "4x GB10, TP4 (spark1-4)",
   "hardware_family": "gb10",
   "units": 4,
   "engine": "vLLM, oxbyte MTP4, EP",
   "quant": "NVFP4",
   "measured": "2026-10-02",
   "metrics": {
    "decode_1u_prose": {
     "value": 41.4,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| Qwen3.8-Flash-Next NVFP4, oxbyte MTP4, EP, 64 seqs | spark1-4, 2300 | 8/8 | 41.4 / 61.1 | 146.7 / 229.7 / 347.5 / 482.4 | 64 (648.7) | 3,571 / 3,530 / n/a (131k ctx) | 439 / 535 | 0.90 (1.18 at 64) |",
     "column": "1 user prose / code",
     "part": 0
    },
    "decode_1u_code": {
     "value": 61.1,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| Qwen3.8-Flash-Next NVFP4, oxbyte MTP4, EP, 64 seqs | spark1-4, 2300 | 8/8 | 41.4 / 61.1 | 146.7 / 229.7 / 347.5 / 482.4 | 64 (648.7) | 3,571 / 3,530 / n/a (131k ctx) | 439 / 535 | 0.90 (1.18 at 64) |",
     "column": "1 user prose / code",
     "part": 1
    },
    "agg_decode_32u": {
     "value": 482.4,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| Qwen3.8-Flash-Next NVFP4, oxbyte MTP4, EP, 64 seqs | spark1-4, 2300 | 8/8 | 41.4 / 61.1 | 146.7 / 229.7 / 347.5 / 482.4 | 64 (648.7) | 3,571 / 3,530 / n/a (131k ctx) | 439 / 535 | 0.90 (1.18 at 64) |",
     "column": "4 / 8 / 16 / 32 users agg (prose)",
     "part": 3
    },
    "ceiling_users": {
     "value": 64,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| Qwen3.8-Flash-Next NVFP4, oxbyte MTP4, EP, 64 seqs | spark1-4, 2300 | 8/8 | 41.4 / 61.1 | 146.7 / 229.7 / 347.5 / 482.4 | 64 (648.7) | 3,571 / 3,530 / n/a (131k ctx) | 439 / 535 | 0.90 (1.18 at 64) |",
     "column": "users at floor (agg)",
     "part": 0
    },
    "agg_decode_at_ceiling": {
     "value": 648.7,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| Qwen3.8-Flash-Next NVFP4, oxbyte MTP4, EP, 64 seqs | spark1-4, 2300 | 8/8 | 41.4 / 61.1 | 146.7 / 229.7 / 347.5 / 482.4 | 64 (648.7) | 3,571 / 3,530 / n/a (131k ctx) | 439 / 535 | 0.90 (1.18 at 64) |",
     "column": "users at floor (agg)",
     "part": 1
    },
    "prefill_8k": {
     "value": 3571,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| Qwen3.8-Flash-Next NVFP4, oxbyte MTP4, EP, 64 seqs | spark1-4, 2300 | 8/8 | 41.4 / 61.1 | 146.7 / 229.7 / 347.5 / 482.4 | 64 (648.7) | 3,571 / 3,530 / n/a (131k ctx) | 439 / 535 | 0.90 (1.18 at 64) |",
     "column": "prefill 8k / 32k / 128k",
     "part": 0
    },
    "prefill_32k": {
     "value": 3530,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| Qwen3.8-Flash-Next NVFP4, oxbyte MTP4, EP, 64 seqs | spark1-4, 2300 | 8/8 | 41.4 / 61.1 | 146.7 / 229.7 / 347.5 / 482.4 | 64 (648.7) | 3,571 / 3,530 / n/a (131k ctx) | 439 / 535 | 0.90 (1.18 at 64) |",
     "column": "prefill 8k / 32k / 128k",
     "part": 1
    },
    "tok_per_joule": {
     "value": 0.9,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| Qwen3.8-Flash-Next NVFP4, oxbyte MTP4, EP, 64 seqs | spark1-4, 2300 | 8/8 | 41.4 / 61.1 | 146.7 / 229.7 / 347.5 / 482.4 | 64 (648.7) | 3,571 / 3,530 / n/a (131k ctx) | 439 / 535 | 0.90 (1.18 at 64) |",
     "column": "tok/J 32 users",
     "part": 0
    }
   },
   "notes": "64 users at the floor, the highest-throughput TP4 layout measured; the MSI copy matched within a few percent."
  },
  {
   "id": "gb10-tp8-qwen3-8-flash-next-nvfp4",
   "model": "Qwen3.8-Flash-Next",
   "model_slug": "qwen3-8-flash-next",
   "hardware": "8x GB10, TP8",
   "hardware_family": "gb10",
   "units": 8,
   "engine": "vLLM, oxbyte MTP4, EP",
   "quant": "NVFP4",
   "measured": "2026-10-03",
   "metrics": {
    "decode_1u_prose": {
     "value": 45.7,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| all 8 at 2300 | 8/8 | 45.7 / 165.5 / 280.2 / 428.2 / 604.7 | 63.5 / 821.3 | 64 (769.7) | 3,736 / 3,672 | 848 / 1,039 | 0.58 / 0.74 | 74 |",
     "column": "1 / 4 / 8 / 16 / 32 users prose agg",
     "part": 0
    },
    "decode_1u_code": {
     "value": 63.5,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| all 8 at 2300 | 8/8 | 45.7 / 165.5 / 280.2 / 428.2 / 604.7 | 63.5 / 821.3 | 64 (769.7) | 3,736 / 3,672 | 848 / 1,039 | 0.58 / 0.74 | 74 |",
     "column": "code 1 / 32",
     "part": 0
    },
    "agg_decode_32u": {
     "value": 604.7,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| all 8 at 2300 | 8/8 | 45.7 / 165.5 / 280.2 / 428.2 / 604.7 | 63.5 / 821.3 | 64 (769.7) | 3,736 / 3,672 | 848 / 1,039 | 0.58 / 0.74 | 74 |",
     "column": "1 / 4 / 8 / 16 / 32 users prose agg",
     "part": 4
    },
    "ceiling_users": {
     "value": 64,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| all 8 at 2300 | 8/8 | 45.7 / 165.5 / 280.2 / 428.2 / 604.7 | 63.5 / 821.3 | 64 (769.7) | 3,736 / 3,672 | 848 / 1,039 | 0.58 / 0.74 | 74 |",
     "column": "users at floor (agg)",
     "part": 0
    },
    "agg_decode_at_ceiling": {
     "value": 769.7,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| all 8 at 2300 | 8/8 | 45.7 / 165.5 / 280.2 / 428.2 / 604.7 | 63.5 / 821.3 | 64 (769.7) | 3,736 / 3,672 | 848 / 1,039 | 0.58 / 0.74 | 74 |",
     "column": "users at floor (agg)",
     "part": 1
    },
    "prefill_8k": {
     "value": 3736,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| all 8 at 2300 | 8/8 | 45.7 / 165.5 / 280.2 / 428.2 / 604.7 | 63.5 / 821.3 | 64 (769.7) | 3,736 / 3,672 | 848 / 1,039 | 0.58 / 0.74 | 74 |",
     "column": "prefill 8k / 32k",
     "part": 0
    },
    "prefill_32k": {
     "value": 3672,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| all 8 at 2300 | 8/8 | 45.7 / 165.5 / 280.2 / 428.2 / 604.7 | 63.5 / 821.3 | 64 (769.7) | 3,736 / 3,672 | 848 / 1,039 | 0.58 / 0.74 | 74 |",
     "column": "prefill 8k / 32k",
     "part": 1
    },
    "tok_per_joule": {
     "value": 0.58,
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "| all 8 at 2300 | 8/8 | 45.7 / 165.5 / 280.2 / 428.2 / 604.7 | 63.5 / 821.3 | 64 (769.7) | 3,736 / 3,672 | 848 / 1,039 | 0.58 / 0.74 | 74 |",
     "column": "tok/J at 32 / at floor",
     "part": 0
    }
   },
   "notes": "One 8-way engine gives 770 tok/s at 64 users against about 1,266 for two TP4 groups of the same model: TP8 only adds single-user speed, matching the 10-01 finding."
  },
  {
   "id": "h200-4x-glm-5-3-flash-fp8-cmmc-rag-v3",
   "model": "GLM-5.3-Flash",
   "model_slug": "glm-5-3-flash",
   "hardware": "4x H200 NVL (ptg-rtx, TP4)",
   "hardware_family": "h200",
   "units": 4,
   "engine": "vLLM",
   "quant": "FP8",
   "measured": "2026-10-02",
   "metrics": {
    "cmmc_rag_strict_pct": {
     "value": 90,
     "unit": "%",
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "GLM-5.3-Flash FP8 agentic strict 90"
    },
    "x_cmmc_rag_frozen_strict": {
     "label": "CMMC RAG v3 frozen-passage strict score, seed 21",
     "value": 83,
     "unit": "count",
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "GLM-5.3-Flash FP8 agentic strict 90, frozen 83"
    },
    "x_cmmc_rag_closed_book": {
     "label": "CMMC RAG v3 closed-book score, seed 21",
     "value": 45,
     "unit": "count",
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "GLM-5.3-Flash FP8 agentic strict 90, frozen 83, closed 45"
    }
   },
   "notes": "Reference rows from the v3 builder (ptg-rtx/H200, seed 21, /92 before item mh_poam_3_10_4)."
  },
  {
   "id": "h200-4x-glm-5-3-flash-fp8-cmmc-rag-v3-ladder",
   "model": "GLM-5.3-Flash",
   "model_slug": "glm-5-3-flash",
   "hardware": "4x H200 NVL (ptg-rtx, TP4)",
   "hardware_family": "h200",
   "units": 4,
   "engine": "vLLM",
   "quant": "FP8",
   "measured": "2026-10-03",
   "metrics": {
    "cmmc_rag_strict_pct": {
     "value": 0.964,
     "unit": "%",
     "source": "fleet-bench/2026-10-03-cmmc-rag-v3-results.md",
     "quote": "| GLM-5.3-Flash FP8 | large MoE | n/a | no | **0.964** (3) | 0.928 (3) | 0.892 (3) | 0.484 (1) | 2.0 -> 0.0 | 2 (24 tok/s/user) or 4 (45) |",
     "column": "Agentic",
     "part": 0
    },
    "x_cmmc_rag_codefirst": {
     "label": "CMMC RAG v3 code-first strict, mean over 3 seeds",
     "value": 0.928,
     "unit": "rate",
     "source": "fleet-bench/2026-10-03-cmmc-rag-v3-results.md",
     "quote": "| GLM-5.3-Flash FP8 | large MoE | n/a | no | **0.964** (3) | 0.928 (3) | 0.892 (3) | 0.484 (1) | 2.0 -> 0.0 | 2 (24 tok/s/user) or 4 (45) |",
     "column": "Code-first",
     "part": 0
    },
    "x_cmmc_rag_frozen": {
     "label": "CMMC RAG v3 frozen strict, mean over 3 seeds",
     "value": 0.892,
     "unit": "rate",
     "source": "fleet-bench/2026-10-03-cmmc-rag-v3-results.md",
     "quote": "| GLM-5.3-Flash FP8 | large MoE | n/a | no | **0.964** (3) | 0.928 (3) | 0.892 (3) | 0.484 (1) | 2.0 -> 0.0 | 2 (24 tok/s/user) or 4 (45) |",
     "column": "Frozen",
     "part": 0
    },
    "x_cmmc_rag_closed": {
     "label": "CMMC RAG v3 closed-book strict, 1 seed",
     "value": 0.484,
     "unit": "rate",
     "source": "fleet-bench/2026-10-03-cmmc-rag-v3-results.md",
     "quote": "| GLM-5.3-Flash FP8 | large MoE | n/a | no | **0.964** (3) | 0.928 (3) | 0.892 (3) | 0.484 (1) | 2.0 -> 0.0 | 2 (24 tok/s/user) or 4 (45) |",
     "column": "Closed",
     "part": 0
    }
   },
   "notes": "CMMC-RAG v3 ladder row; GLM-5.3-Flash FP8 served from the ptg-rtx 4x H200 production endpoint."
  },
  {
   "id": "h200-4x-glm-5-3-flash-fp8-tp4-mtp3",
   "model": "GLM-5.3-Flash",
   "model_slug": "glm-5-3-flash",
   "hardware": "4x H200 NVL (ptg-rtx, TP4)",
   "hardware_family": "h200",
   "units": 4,
   "engine": "vLLM",
   "quant": "FP8",
   "measured": "2026-09-25",
   "metrics": {
    "decode_1u_prose": {
     "value": 238,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-25-dgx1-best-model-eval.md",
     "quote": "| 1 user | 238 tok/s (MTP3); 144 without MTP | 238 tok/s |",
     "column": "4x H200 FP8"
    },
    "agg_decode_32u": {
     "value": 2254,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-25-dgx1-best-model-eval.md",
     "quote": "| 32 users | 1,713 aggregate (MTP3); 1,633 without MTP (1 run) | 2,254 aggregate |",
     "column": "4x H200 FP8"
    },
    "peak_agg_decode": {
     "value": 5245,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-25-dgx1-best-model-eval.md",
     "quote": "5,245 at 256 users (32 tok/s each)"
    },
    "prefill_8k": {
     "value": 20763.5,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-25-dgx1-eval-glm53-flash-fp8-h200.json",
     "quote": "urrency\": 1, \"rep\": 0, \"max_tokens\": 32, \"requests\": 1, \"ok\": 1, \"failed\": 0, \"prompt_tokens_each\": [ 8192 ], \"prompt_tokens_total\": 8192, \"ttft_p50_s\": 0.394, \"ttft_p95_s\": 0.394, \"ttft_min_s\": 0.394, \"ttft_max_s\": 0.394, \"time_to_last_first_token_s\": 0.395, \"aggregate_prefill_tok_s\": 20763.5, \"decode_tok_s_per_stream_p50\": 241.4, ",
     "json_key": "aggregate_prefill_tok_s"
    },
    "prefill_32k": {
     "value": 22696.7,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-25-dgx1-eval-glm53-flash-fp8-h200.json",
     "quote": "rency\": 1, \"rep\": 0, \"max_tokens\": 32, \"requests\": 1, \"ok\": 1, \"failed\": 0, \"prompt_tokens_each\": [ 32768 ], \"prompt_tokens_total\": 32768, \"ttft_p50_s\": 1.444, \"ttft_p95_s\": 1.444, \"ttft_min_s\": 1.444, \"ttft_max_s\": 1.444, \"time_to_last_first_token_s\": 1.444, \"aggregate_prefill_tok_s\": 22696.7, \"decode_tok_s_per_stream_p50\": 342.6, ",
     "json_key": "aggregate_prefill_tok_s"
    },
    "prefill_128k": {
     "value": 21999.8,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-25-dgx1-eval-glm53-flash-fp8-h200.json",
     "quote": "ncy\": 1, \"rep\": 0, \"max_tokens\": 32, \"requests\": 1, \"ok\": 1, \"failed\": 0, \"prompt_tokens_each\": [ 131040 ], \"prompt_tokens_total\": 131040, \"ttft_p50_s\": 5.956, \"ttft_p95_s\": 5.956, \"ttft_min_s\": 5.956, \"ttft_max_s\": 5.956, \"time_to_last_first_token_s\": 5.956, \"aggregate_prefill_tok_s\": 21999.8, \"decode_tok_s_per_stream_p50\": 297.4, ",
     "json_key": "aggregate_prefill_tok_s"
    },
    "ttft_32k": {
     "value": 1.444,
     "unit": "s",
     "source": "fleet-bench/2026-09-25-dgx1-eval-glm53-flash-fp8-h200.json",
     "quote": "rency\": 1, \"rep\": 0, \"max_tokens\": 32, \"requests\": 1, \"ok\": 1, \"failed\": 0, \"prompt_tokens_each\": [ 32768 ], \"prompt_tokens_total\": 32768, \"ttft_p50_s\": 1.444, \"ttft_p95_s\": 1.444, \"ttft_min_s\": 1.444, \"ttft_max_s\": 1.444, \"time_to_last_first_token_s\": 1.444, \"aggregate_prefill_t",
     "json_key": "ttft_p50_s"
    },
    "context_verified": {
     "value": 510997,
     "unit": "tokens",
     "source": "fleet-bench/2026-09-25-dgx1-best-model-eval.md",
     "quote": "524,288. A needle at 510,997 tokens passed (27.8 s end to end)"
    },
    "toolcall_pass": {
     "value": 23,
     "unit": "count",
     "source": "fleet-bench/2026-09-25-dgx1-eval-glm53-flash-fp8-h200.json",
     "quote": "_tokens\": 12709 } }, \"write_judges_overall_mean\": { \"ocloud/kimi-k3\": 9 } }, \"llmbench_tool_calls_v2\": { \"n_cases\": 27, \"n_passed\": 23, \"api_errors\": 0, \"wall_clock_seconds\": 62.7,",
     "json_key": "n_passed"
    },
    "x_toolcall_accuracy_pct": {
     "label": "Tool-call v2 overall accuracy",
     "value": 85.2,
     "unit": "%",
     "source": "fleet-bench/2026-09-25-dgx1-eval-glm53-flash-fp8-h200.json",
     "quote": "ll_clock_seconds\": 62.7, \"scorecard\": { \"overall_accuracy_pct\": 85.2, \"multi_turn_calendar_pct\": 100.0, \"multi_turn_crm_",
     "json_key": "overall_accuracy_pct"
    },
    "x_agentic_completion_rate": {
     "label": "Agentic tool completion rate (12 cases)",
     "value": 0.917,
     "unit": "rate",
     "source": "fleet-bench/2026-09-25-dgx1-eval-glm53-flash-fp8-h200.json",
     "quote": "mbench_agentic_tools\": { \"n_cases\": 12, \"completion_rate\": 0.917, \"fabrication_rate\": 0.083, \"honest_incomplete_rate\": 0",
     "json_key": "completion_rate"
    },
    "x_hallucination_rate_pct": {
     "label": "Tool-call hallucination rate",
     "value": 0.0,
     "unit": "%",
     "source": "fleet-bench/2026-09-25-dgx1-eval-glm53-flash-fp8-h200.json",
     "quote": "hain_pct\": 50.0, \"negative_pct\": 100.0, \"hallucination_rate_pct\": 0.0, \"order_vi",
     "json_key": "hallucination_rate_pct"
    },
    "x_kv_cache_tokens": {
     "label": "KV cache capacity from /metrics",
     "value": 3329280,
     "unit": "tokens",
     "source": "fleet-bench/2026-09-25-dgx1-eval-glm53-flash-fp8-h200.json",
     "quote": "\"kv_cache_tokens_from_metrics\": 3329280 }",
     "json_key": "kv_cache_tokens_from_metrics"
    },
    "x_evalsuite_total": {
     "label": "Eval suite total passed (tools 38, agent 15, cmmc 28, code 10, write 3, long 26)",
     "value": 120,
     "unit": "count",
     "source": "fleet-bench/2026-09-25-dgx1-best-model-eval.md",
     "quote": "| GLM-5.3-Flash FP8, 4x H200 | 38/38 | 15/15 | 28/35 (5-run mean 27.6) | 10/10 | 3/3 | 26/26 | 120/127 | 85.2% (0% halluc.) | 0.917 / 0.083, 6.0 turns | 363 s (conc 8) |",
     "column": "total",
     "part": 0
    }
   },
   "notes": "Production service llm-glm53-flash on ptg-rtx (zai-org/GLM-5.3-Flash FP8, TP4 on 4x H200 NVL, MTP x3, max-model-len 524288), measured read-only during the dgx1 best-model eval."
  },
  {
   "id": "h200-4x-gpt-oss-120b-cmmc-rag-v3",
   "model": "gpt-oss-120b",
   "model_slug": "gpt-oss-120b",
   "hardware": "4x H200 NVL (ptg-rtx, TP4)",
   "hardware_family": "h200",
   "units": 4,
   "engine": "vLLM",
   "quant": "FP8",
   "measured": "2026-10-02",
   "metrics": {
    "cmmc_rag_strict_pct": {
     "value": 61,
     "unit": "%",
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "gpt-oss-120b agentic 61"
    },
    "x_cmmc_rag_closed_book": {
     "label": "CMMC RAG v3 closed-book score, seed 21",
     "value": 30,
     "unit": "count",
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "gpt-oss-120b agentic 61, closed 30"
    }
   },
   "notes": "Reference row from the v3 builder (ptg-rtx/H200, seed 21)."
  },
  {
   "id": "h200-4x-qwen3-8-27b-fp8-cmmc-rag-v3",
   "model": "Qwen3.8-27B (FP8)",
   "model_slug": "qwen3-8-27b",
   "hardware": "4x H200 NVL (ptg-rtx, TP4)",
   "hardware_family": "h200",
   "units": 4,
   "engine": "vLLM",
   "quant": "FP8",
   "measured": "2026-10-02",
   "metrics": {
    "cmmc_rag_strict_pct": {
     "value": 90,
     "unit": "%",
     "source": "fleet-bench/2026-10-02-best-model-per-tier-results.md",
     "quote": "Qwen3.8-27B FP8 agentic 90"
    }
   },
   "notes": "Reference row from the v3 builder (ptg-rtx/H200, seed 21)."
  },
  {
   "id": "rtxpro-1x-gpt-oss-120b-mxfp4-cmmc-rag-v3",
   "model": "gpt-oss-120b",
   "model_slug": "gpt-oss-120b",
   "hardware": "1x RTX PRO 6000 Blackwell Max-Q (h200monsta)",
   "hardware_family": "rtx-pro-6000",
   "units": 1,
   "engine": "vLLM",
   "quant": "MXFP4",
   "measured": "2026-10-03",
   "metrics": {
    "cmmc_rag_strict_pct": {
     "value": 0.76,
     "unit": "%",
     "source": "fleet-bench/2026-10-03-cmmc-rag-v3-results.md",
     "quote": "| gpt-oss-120b MXFP4 (h200monsta) | 117 (5.1) | 61 | by memory, CPU-slow | 0.760 (3) | 0.889 (3) | 0.860 (3) | 0.333 (1) | 1.7 -> 0.0 | 1 |",
     "column": "Agentic",
     "part": 0
    },
    "x_cmmc_rag_codefirst": {
     "label": "CMMC RAG v3 code-first strict, mean over 3 seeds",
     "value": 0.889,
     "unit": "rate",
     "source": "fleet-bench/2026-10-03-cmmc-rag-v3-results.md",
     "quote": "| gpt-oss-120b MXFP4 (h200monsta) | 117 (5.1) | 61 | by memory, CPU-slow | 0.760 (3) | 0.889 (3) | 0.860 (3) | 0.333 (1) | 1.7 -> 0.0 | 1 |",
     "column": "Code-first",
     "part": 0
    },
    "x_cmmc_rag_frozen": {
     "label": "CMMC RAG v3 frozen strict, mean over 3 seeds",
     "value": 0.86,
     "unit": "rate",
     "source": "fleet-bench/2026-10-03-cmmc-rag-v3-results.md",
     "quote": "| gpt-oss-120b MXFP4 (h200monsta) | 117 (5.1) | 61 | by memory, CPU-slow | 0.760 (3) | 0.889 (3) | 0.860 (3) | 0.333 (1) | 1.7 -> 0.0 | 1 |",
     "column": "Frozen",
     "part": 0
    },
    "x_cmmc_rag_closed": {
     "label": "CMMC RAG v3 closed-book strict, 1 seed",
     "value": 0.333,
     "unit": "rate",
     "source": "fleet-bench/2026-10-03-cmmc-rag-v3-results.md",
     "quote": "| gpt-oss-120b MXFP4 (h200monsta) | 117 (5.1) | 61 | by memory, CPU-slow | 0.760 (3) | 0.889 (3) | 0.860 (3) | 0.333 (1) | 1.7 -> 0.0 | 1 |",
     "column": "Closed",
     "part": 0
    }
   },
   "notes": "CMMC-RAG v3 ladder row; gpt-oss-120b MXFP4 on h200monsta."
  },
  {
   "id": "rtxpro-1x-gpt-oss-120b-mxfp4-tp1",
   "model": "gpt-oss-120b",
   "model_slug": "gpt-oss-120b",
   "hardware": "1x RTX PRO 6000 Blackwell Max-Q (h200monsta)",
   "hardware_family": "rtx-pro-6000",
   "units": 1,
   "engine": "vLLM",
   "quant": "MXFP4",
   "measured": "2026-09-21",
   "metrics": {
    "decode_1u_prose": {
     "value": 179.0,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-21-h200monsta-gpt-oss-120b-judge-shape.json",
     "quote": "{ \"base\": \"http://h200monsta:9402/v1\", \"model\": \"gpt-oss-120b\", \"max_tokens\": 1700, \"pad_words\": 3000, \"thinking\": false, \"measured_at\": \"2026-09-21T12:57:50-0400\", \"rows\": [ { \"concurrency\": 1, \"ok\": 1, \"failed\": 0, \"wall_s\": 7.52, \"aggregate_tok_s\": 179.0, \"prompt_tokens_median\": 3097, \"request",
     "json_key": "aggregate_tok_s"
    },
    "x_per_stream_1u": {
     "label": "Per-stream decode median, 1 user",
     "value": 203.5,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-21-h200monsta-gpt-oss-120b-judge-shape.json",
     "quote": "ss-120b\", \"max_tokens\": 1700, \"pad_words\": 3000, \"thinking\": false, \"measured_at\": \"2026-09-21T12:57:50-0400\", \"rows\": [ { \"concurrency\": 1, \"ok\": 1, \"failed\": 0, \"wall_s\": 7.52, \"aggregate_tok_s\": 179.0, \"prompt_tokens_median\": 3097, \"requests_per_min\": 8.0, \"per_stream_tok_s_median\": 203.5, \"ttft_p50_s\": 0.904, \"ttft_p95_s\": 0.9",
     "json_key": "per_stream_tok_s_median"
    },
    "agg_decode_8u": {
     "value": 570.0,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-21-h200monsta-gpt-oss-120b-judge-shape.json",
     "quote": "\"errors\": [] }, { \"concurrency\": 8, \"ok\": 8, \"failed\": 0, \"wall_s\": 20.81, \"aggregate_tok_s\": 570.0, \"prompt_tokens_median\": 3098.0",
     "json_key": "aggregate_tok_s"
    },
    "agg_decode_32u": {
     "value": 1118.1,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-21-h200monsta-gpt-oss-120b-judge-shape.json",
     "quote": "\"errors\": [] }, { \"concurrency\": 32, \"ok\": 32, \"failed\": 0, \"wall_s\": 44.51, \"aggregate_tok_s\": 1118.1, \"prompt_tokens_median\": 3098.0",
     "json_key": "aggregate_tok_s"
    },
    "peak_agg_decode": {
     "value": 1511.2,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-21-h200monsta-gpt-oss-120b-judge-shape.json",
     "quote": "\"errors\": [] }, { \"concurrency\": 64, \"ok\": 64, \"failed\": 0, \"wall_s\": 63.2, \"aggregate_tok_s\": 1511.2, \"prompt_tokens_median\": 3098.0",
     "json_key": "aggregate_tok_s"
    },
    "ceiling_users": {
     "value": 64,
     "unit": "users",
     "source": "fleet-bench/2026-09-21-h200monsta-gpt-oss-120b-judge-shape.json",
     "quote": "\"errors\": [] }, { \"concurrency\": 64, \"ok\": 64, \"failed\": 0, \"wall_s\": 63.2, \"aggregate_tok_s\": 1511.2",
     "json_key": "concurrency"
    },
    "x_ttft_1u_s": {
     "label": "TTFT p50, 1 user (3,097-token SEO-judge prompt)",
     "value": 0.904,
     "unit": "s",
     "source": "fleet-bench/2026-09-21-h200monsta-gpt-oss-120b-judge-shape.json",
     "quote": "T12:57:50-0400\", \"rows\": [ { \"concurrency\": 1, \"ok\": 1, \"failed\": 0, \"wall_s\": 7.52, \"aggregate_tok_s\": 179.0, \"prompt_tokens_median\": 3097, \"requests_per_min\": 8.0, \"per_stream_tok_s_median\": 203.5, \"ttft_p50_s\": 0.904,",
     "json_key": "ttft_p50_s"
    },
    "x_ttft_32u_s": {
     "label": "TTFT p50, 32 users",
     "value": 7.99,
     "unit": "s",
     "source": "fleet-bench/2026-09-21-h200monsta-gpt-oss-120b-judge-shape.json",
     "quote": "\"errors\": [] }, { \"concurrency\": 32, \"ok\": 32, \"failed\": 0, \"wall_s\": 44.51, \"aggregate_tok_s\": 1118.1, \"prompt_tokens_median\": 3098.0, \"requests_per_min\": 43.1, \"per_stream_tok_s_median\": 46.9, \"ttft_p50_s\": 7.99",
     "json_key": "ttft_p50_s"
    },
    "x_prompt_tokens_median": {
     "label": "Prompt tokens, median (padded judge prompt)",
     "value": 3098,
     "unit": "tokens",
     "source": "fleet-bench/2026-09-21-h200monsta-gpt-oss-120b-judge-shape.json",
     "quote": "\"errors\": [] }, { \"concurrency\": 8, \"ok\": 8, \"failed\": 0, \"wall_s\": 20.81, \"aggregate_tok_s\": 570.0, \"prompt_tokens_median\": 3098.0",
     "json_key": "prompt_tokens_median"
    }
   },
   "notes": "SEO judge shape workload on the live gpt-oss-120b judge service (:9402, GPU 1), 1,700 max output tokens, no thinking."
  },
  {
   "id": "rtxpro-1x-qwen3-8-27b-fp8-cmmc-rag-v3",
   "model": "Qwen3.8-27B (FP8)",
   "model_slug": "qwen3-8-27b",
   "hardware": "1x RTX PRO 6000 Blackwell Max-Q (h200monsta)",
   "hardware_family": "rtx-pro-6000",
   "units": 1,
   "engine": "vLLM",
   "quant": "FP8",
   "measured": "2026-10-03",
   "metrics": {
    "cmmc_rag_strict_pct": {
     "value": 0.971,
     "unit": "%",
     "source": "fleet-bench/2026-10-03-cmmc-rag-v3-results.md",
     "quote": "| Qwen3.8-27B FP8 (h200monsta) | 27.3 dense | 29 | no | **0.971** (3) | 0.925 (3) | 0.875 (3) | 0.258 (1) | 1.3 -> 0.0 | 1 |",
     "column": "Agentic",
     "part": 0
    },
    "x_cmmc_rag_codefirst": {
     "label": "CMMC RAG v3 code-first strict, mean over 3 seeds",
     "value": 0.925,
     "unit": "rate",
     "source": "fleet-bench/2026-10-03-cmmc-rag-v3-results.md",
     "quote": "| Qwen3.8-27B FP8 (h200monsta) | 27.3 dense | 29 | no | **0.971** (3) | 0.925 (3) | 0.875 (3) | 0.258 (1) | 1.3 -> 0.0 | 1 |",
     "column": "Code-first",
     "part": 0
    },
    "x_cmmc_rag_frozen": {
     "label": "CMMC RAG v3 frozen strict, mean over 3 seeds",
     "value": 0.875,
     "unit": "rate",
     "source": "fleet-bench/2026-10-03-cmmc-rag-v3-results.md",
     "quote": "| Qwen3.8-27B FP8 (h200monsta) | 27.3 dense | 29 | no | **0.971** (3) | 0.925 (3) | 0.875 (3) | 0.258 (1) | 1.3 -> 0.0 | 1 |",
     "column": "Frozen",
     "part": 0
    },
    "x_cmmc_rag_closed": {
     "label": "CMMC RAG v3 closed-book strict, 1 seed",
     "value": 0.258,
     "unit": "rate",
     "source": "fleet-bench/2026-10-03-cmmc-rag-v3-results.md",
     "quote": "| Qwen3.8-27B FP8 (h200monsta) | 27.3 dense | 29 | no | **0.971** (3) | 0.925 (3) | 0.875 (3) | 0.258 (1) | 1.3 -> 0.0 | 1 |",
     "column": "Closed",
     "part": 0
    }
   },
   "notes": "CMMC-RAG v3 ladder row; Qwen3.8-27B FP8 on h200monsta GPU 0."
  },
  {
   "id": "rtxpro-1x-qwen3-8-27b-fp8-seo-shape-seqs32",
   "model": "Qwen3.8-27B (FP8)",
   "model_slug": "qwen3-8-27b",
   "hardware": "1x RTX PRO 6000 Blackwell Max-Q (h200monsta)",
   "hardware_family": "rtx-pro-6000",
   "units": 1,
   "engine": "vLLM",
   "quant": "FP8",
   "measured": "2026-09-21",
   "metrics": {
    "agg_decode_8u": {
     "value": 264.4,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-21-h200monsta-qwen38-27b-seo-shape-seqs32.json",
     "quote": "\"measured_at\": \"2026-09-21T12:37:47-0400\", \"rows\": [ { \"concurrency\": 8, \"ok\": 8, \"failed\": 0, \"wall_s\": 40.51, \"aggregate_tok_s\": 264.4, \"prompt_tokens_median\": 8645",
     "json_key": "aggregate_tok_s"
    },
    "agg_decode_32u": {
     "value": 331.6,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-21-h200monsta-qwen38-27b-seo-shape-seqs32.json",
     "quote": "\"errors\": [] }, { \"concurrency\": 32, \"ok\": 32, \"failed\": 0, \"wall_s\": 106.68, \"aggregate_tok_s\": 331.6, \"prompt_tokens_median\": 8646",
     "json_key": "aggregate_tok_s"
    },
    "x_ttft_8u_s": {
     "label": "TTFT p50, 8 users (8,645-token SEO prompt)",
     "value": 11.447,
     "unit": "s",
     "source": "fleet-bench/2026-09-21-h200monsta-qwen38-27b-seo-shape-seqs32.json",
     "quote": "rds\": 8600, \"thinking\": false, \"measured_at\": \"2026-09-21T12:37:47-0400\", \"rows\": [ { \"concurrency\": 8, \"ok\": 8, \"failed\": 0, \"wall_s\": 40.51, \"aggregate_tok_s\": 264.4, \"prompt_tokens_median\": 8645.0, \"requests_per_min\": 11.8, \"per_stream_tok_s_median\": 48.1, \"ttft_p50_s\": 11.447, \"ttft_p95_s\": 13.4",
     "json_key": "ttft_p50_s"
    },
    "x_prompt_tokens_median": {
     "label": "Prompt tokens, median (padded SEO shape)",
     "value": 8646,
     "unit": "tokens",
     "source": "fleet-bench/2026-09-21-h200monsta-qwen38-27b-seo-shape-seqs32.json",
     "quote": "\"errors\": [] }, { \"concurrency\": 16, \"ok\": 16, \"failed\": 0, \"wall_s\": 60.44, \"aggregate_tok_s\": 301.2, \"prompt_tokens_median\": 8646.0",
     "json_key": "prompt_tokens_median"
    }
   },
   "notes": "SEO-shape workload (8,600-word padded prompt, 1,500 max output tokens, thinking off) on the live qwen3.8-27b-fp8 service, max_num_seqs 32. Long-prompt aggregate throughput, not comparable to the small-prompt decode ladder."
  },
  {
   "id": "rtxpro-1x-qwen3-8-27b-fp8-seo-shape-seqs64-bt32768",
   "model": "Qwen3.8-27B (FP8)",
   "model_slug": "qwen3-8-27b",
   "hardware": "1x RTX PRO 6000 Blackwell Max-Q (h200monsta)",
   "hardware_family": "rtx-pro-6000",
   "units": 1,
   "engine": "vLLM",
   "quant": "FP8",
   "measured": "2026-09-21",
   "metrics": {
    "agg_decode_32u": {
     "value": 259.5,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-21-h200monsta-qwen38-27b-seo-shape-seqs64-bt32768.json",
     "quote": "\"errors\": [] }, { \"concurrency\": 32, \"ok\": 32, \"failed\": 0, \"wall_s\": 151.86, \"aggregate_tok_s\": 259.5, \"prompt_tokens_median\": 8646",
     "json_key": "aggregate_tok_s"
    },
    "peak_agg_decode": {
     "value": 270.9,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-21-h200monsta-qwen38-27b-seo-shape-seqs64-bt32768.json",
     "quote": "\"errors\": [] }, { \"concurrency\": 64, \"ok\": 64, \"failed\": 0, \"wall_s\": 278.3, \"aggregate_tok_s\": 270.9, \"prompt_tokens_median\": 8647",
     "json_key": "aggregate_tok_s"
    },
    "ceiling_users": {
     "value": 64,
     "unit": "users",
     "source": "fleet-bench/2026-09-21-h200monsta-qwen38-27b-seo-shape-seqs64-bt32768.json",
     "quote": "\"errors\": [] }, { \"concurrency\": 64, \"ok\": 64, \"failed\": 0, \"wall_s\": 278.3, \"aggregate_tok_s\": 270.9",
     "json_key": "concurrency"
    },
    "x_ttft_16u_s": {
     "label": "TTFT p50, 16 users",
     "value": 46.956,
     "unit": "s",
     "source": "fleet-bench/2026-09-21-h200monsta-qwen38-27b-seo-shape-seqs64-bt32768.json",
     "quote": ": false, \"measured_at\": \"2026-09-21T12:49:17-0400\", \"rows\": [ { \"concurrency\": 16, \"ok\": 16, \"failed\": 0, \"wall_s\": 83.62, \"aggregate_tok_s\": 230.0, \"prompt_tokens_median\": 8646.0, \"requests_per_min\": 11.5, \"per_stream_tok_s_median\": 39.3, \"ttft_p50_s\": 46.956, \"ttft_p95_s\": 48.3",
     "json_key": "ttft_p50_s"
    },
    "x_agg_decode_16u": {
     "value": 230.0,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-21-h200monsta-qwen38-27b-seo-shape-seqs64-bt32768.json",
     "quote": "{ \"base\": \"http://h200monsta:9401/v1\", \"model\": \"qwen3.8-27b-fp8\", \"max_tokens\": 1500, \"pad_words\": 8600, \"thinking\": false, \"measured_at\": \"2026-09-21T12:49:17-0400\", \"rows\": [ { \"concurrency\": 16, \"ok\": 16, \"failed\": 0, \"wall_s\": 83.62, \"aggregate_tok_s\": 230.0, \"prompt_tokens_median\": 8646",
     "label": "Aggregate decode at 16 users",
     "json_key": "aggregate_tok_s"
    }
   },
   "notes": "SEO-shape workload with max_num_seqs 64 and block size 32768: higher concurrency than the seqs32 run but slower per request."
  },
  {
   "id": "rtxpro-1x-qwen3-8-27b-fp8-tp1",
   "model": "Qwen3.8-27B (FP8)",
   "model_slug": "qwen3-8-27b",
   "hardware": "1x RTX PRO 6000 Blackwell Max-Q (h200monsta)",
   "hardware_family": "rtx-pro-6000",
   "units": 1,
   "engine": "vLLM",
   "quant": "FP8",
   "measured": "2026-09-23",
   "metrics": {
    "decode_1u_prose": {
     "value": 80.1,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-23-h200monsta-qwen38-27b-concurrency.json",
     "quote": "{ \"base\": \"http://h200monsta:9401/v1\", \"model\": \"qwen3.8-27b-fp8\", \"max_tokens\": 256, \"pad_words\": 0, \"thinking\": false, \"measured_at\": \"2026-09-23T15:50:11-0400\", \"rows\": [ { \"concurrency\": 1, \"ok\": 1, \"failed\": 0, \"wall_s\": 3.2, \"aggregate_tok_s\": 80.1, \"prompt_tokens_median\": 39, \"requests_",
     "json_key": "aggregate_tok_s"
    },
    "agg_decode_8u": {
     "value": 568.8,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-23-h200monsta-qwen38-27b-concurrency.json",
     "quote": "\"errors\": [] }, { \"concurrency\": 8, \"ok\": 8, \"failed\": 0, \"wall_s\": 3.6, \"aggregate_tok_s\": 568.8, \"prompt_tokens_median\": 40.0, \"request",
     "json_key": "aggregate_tok_s"
    },
    "peak_agg_decode": {
     "value": 1358.0,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-23-h200monsta-qwen38-27b-concurrency.json",
     "quote": "\"errors\": [] }, { \"concurrency\": 32, \"ok\": 32, \"failed\": 0, \"wall_s\": 6.03, \"aggregate_tok_s\": 1358.0, \"prompt_tokens_median\": 41.0, \"request",
     "json_key": "aggregate_tok_s"
    },
    "ceiling_users": {
     "value": 32,
     "unit": "users",
     "source": "fleet-bench/2026-09-23-h200monsta-qwen38-27b-concurrency.json",
     "quote": "\"errors\": [] }, { \"concurrency\": 32, \"ok\": 32, \"failed\": 0, \"wall_s\": 6.03, \"aggregate_tok_s\": 1358.0",
     "json_key": "concurrency"
    },
    "x_agg_decode_16u": {
     "label": "Aggregate decode at 16 users",
     "value": 943.2,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-23-h200monsta-qwen38-27b-concurrency.json",
     "quote": "\"errors\": [] }, { \"concurrency\": 16, \"ok\": 16, \"failed\": 0, \"wall_s\": 4.34, \"aggregate_tok_s\": 943.2, \"prompt_tokens_median\": 41.0",
     "json_key": "aggregate_tok_s"
    },
    "x_agg_decode_48u": {
     "label": "Aggregate decode at 48 users (queued past max_num_seqs 32)",
     "value": 1190.2,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-23-h200monsta-qwen38-27b-concurrency.json",
     "quote": "\"errors\": [] }, { \"concurrency\": 48, \"ok\": 48, \"failed\": 0, \"wall_s\": 10.32, \"aggregate_tok_s\": 1190.2, \"prompt_tokens_median\": 42.0",
     "json_key": "aggregate_tok_s"
    },
    "x_per_stream_1u": {
     "label": "Per-stream decode median, 1 user",
     "value": 82.9,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-23-h200monsta-qwen38-27b-concurrency.json",
     "quote": "wen3.8-27b-fp8\", \"max_tokens\": 256, \"pad_words\": 0, \"thinking\": false, \"measured_at\": \"2026-09-23T15:50:11-0400\", \"rows\": [ { \"concurrency\": 1, \"ok\": 1, \"failed\": 0, \"wall_s\": 3.2, \"aggregate_tok_s\": 80.1, \"prompt_tokens_median\": 39, \"requests_per_min\": 18.8, \"per_stream_tok_s_median\": 82.9, \"ttft_p50_s\": 0.108, \"ttft_p95_s\": 0.1",
     "json_key": "per_stream_tok_s_median"
    },
    "context_verified": {
     "value": 131072,
     "unit": "tokens",
     "source": "fleet-bench/2026-09-23-h200monsta-qwen38-27b-concurrency.json",
     "quote": "max-model-len 131072"
    }
   },
   "notes": "Re-measurement of the 2026-09-21 run on the same live vLLM service (TP=1 on GPU 0, max-num-seqs 32, MTP spec-decode 2, prefix caching). Supersedes the 09-21 run (values reproduced within noise). The other three GPUs were busy, so this is one RTX PRO 6000 card, not four."
  },
  {
   "id": "rtxpro-1x-rtx-pro-6000-maxq-baseline",
   "model": "Hardware baseline (PyTorch, no LLM)",
   "model_slug": "rtx-pro-6000-maxq-baseline",
   "hardware": "1x RTX PRO 6000 Blackwell Max-Q (h200monsta)",
   "hardware_family": "rtx-pro-6000",
   "units": 1,
   "engine": "PyTorch 2.13.0+cu130",
   "quant": "BF16",
   "measured": "2026-09-23",
   "metrics": {
    "x_copy_bandwidth": {
     "label": "Single-GPU device copy bandwidth (best)",
     "value": 1468.0,
     "unit": "GB/s",
     "source": "fleet-bench/2026-09-23-h200monsta-gpu-hardware-baseline.json",
     "quote": "88, \"vram_total_gib\": 95.0, \"vram_type\": \"GDDR7 (not HBM)\", \"copy_bandwidth_gb_s_best\": 1468.0, \"copy_bandwidth_gb_s_median\": 1464.3, ",
     "json_key": "copy_bandwidth_gb_s_best"
    },
    "x_read_bandwidth": {
     "label": "Single-GPU device read bandwidth (best)",
     "value": 1633.1,
     "unit": "GB/s",
     "source": "fleet-bench/2026-09-23-h200monsta-gpu-hardware-baseline.json",
     "quote": "h_gb_s_best\": 1468.0, \"copy_bandwidth_gb_s_median\": 1464.3, \"read_bandwidth_gb_s_best\": 1633.1, \"bf16_matmul_tflop_s_best_by_n\": { \"40",
     "json_key": "read_bandwidth_gb_s_best"
    },
    "power_w": {
     "value": 300,
     "unit": "W",
     "source": "fleet-bench/2026-09-23-h200monsta-gpu-hardware-baseline.json",
     "quote": "\"variant\": \"Max-Q\", \"power_limit_w\": 300, \"compute_capabilit",
     "json_key": "power_limit_w"
    }
   },
   "notes": "Clean single-GPU hardware baseline on idle GPU 2 (RTX PRO 6000 Blackwell Max-Q). The h200monsta name is historical: the box holds 4x RTX PRO 6000 Blackwell, no H200 cards (record correction in the source)."
  },
  {
   "id": "rtxpro-2x-glm-5-3-flash-fp8-tp2-aligned",
   "model": "GLM-5.3-Flash",
   "model_slug": "glm-5-3-flash",
   "hardware": "2x RTX PRO 6000 Blackwell Max-Q (h200monsta)",
   "hardware_family": "rtx-pro-6000",
   "units": 2,
   "engine": "vLLM",
   "quant": "FP8",
   "measured": "2026-10-10",
   "metrics": {
    "decode_1u_prose": {
     "value": 166.0,
     "unit": "tok/s",
     "source": "fleet-bench/h200monsta-glm-recipe-2026-10-10/summary.md",
     "quote": "| aligned | 190 | 0.32 | 0.37 | 0.53 | 166 | 394 | 430 | 0 | 224 | 0 | 2 | 0.74 | 291/415 |",
     "column": "decode tok/s"
    },
    "x_sam_first_token_s": {
     "label": "Sam first-token latency, new turn",
     "value": 0.37,
     "unit": "s",
     "source": "fleet-bench/h200monsta-glm-recipe-2026-10-10/summary.md",
     "quote": "| aligned | 190 | 0.32 | 0.37 | 0.53 | 166 | 394 | 430 | 0 | 224 | 0 | 2 | 0.74 | 291/415 |",
     "column": "Sam 1st tok, new turn"
    },
    "x_sam_turn_total_s": {
     "label": "Sam turn total, median",
     "value": 0.53,
     "unit": "s",
     "source": "fleet-bench/h200monsta-glm-recipe-2026-10-10/summary.md",
     "quote": "| aligned | 190 | 0.32 | 0.37 | 0.53 | 166 | 394 | 430 | 0 | 224 | 0 | 2 | 0.74 | 291/415 |",
     "column": "Sam turn total med"
    },
    "x_agg_decode_8stream": {
     "label": "Aggregate decode at 8 streams",
     "value": 394.0,
     "unit": "tok/s",
     "source": "fleet-bench/h200monsta-glm-recipe-2026-10-10/summary.md",
     "quote": "| aligned | 190 | 0.32 | 0.37 | 0.53 | 166 | 394 | 430 | 0 | 224 | 0 | 2 | 0.74 | 291/415 |",
     "column": "8-stream tok/s"
    },
    "x_agg_decode_16stream": {
     "label": "Aggregate decode at 16 streams",
     "value": 430.0,
     "unit": "tok/s",
     "source": "fleet-bench/h200monsta-glm-recipe-2026-10-10/summary.md",
     "quote": "| aligned | 190 | 0.32 | 0.37 | 0.53 | 166 | 394 | 430 | 0 | 224 | 0 | 2 | 0.74 | 291/415 |",
     "column": "16-stream tok/s"
    },
    "x_corpus_score": {
     "label": "GLM corpus score of 264",
     "value": 224,
     "unit": "count",
     "source": "fleet-bench/h200monsta-glm-recipe-2026-10-10/summary.md",
     "quote": "| aligned | 190 | 0.32 | 0.37 | 0.53 | 166 | 394 | 430 | 0 | 224 | 0 | 2 | 0.74 | 291/415 |",
     "column": "corpus /264"
    }
   },
   "notes": "TP2 recipe sweep config 'aligned' on h200monsta GPUs 1,3 (:9420). Decodes are single-stream Sam-turn measurements."
  },
  {
   "id": "rtxpro-2x-glm-5-3-flash-fp8-tp2-aligned-all",
   "model": "GLM-5.3-Flash",
   "model_slug": "glm-5-3-flash",
   "hardware": "2x RTX PRO 6000 Blackwell Max-Q (h200monsta)",
   "hardware_family": "rtx-pro-6000",
   "units": 2,
   "engine": "vLLM",
   "quant": "FP8",
   "measured": "2026-10-10",
   "metrics": {
    "decode_1u_prose": {
     "value": 166.0,
     "unit": "tok/s",
     "source": "fleet-bench/h200monsta-glm-recipe-2026-10-10/summary.md",
     "quote": "| aligned-all | 190 | 0.31 | 0.34 | 0.47 | 166 | 444 | 419 | 0 | 221 | 1 | 2 | 0.72 | 302/442 |",
     "column": "decode tok/s"
    },
    "x_sam_first_token_s": {
     "label": "Sam first-token latency, new turn",
     "value": 0.34,
     "unit": "s",
     "source": "fleet-bench/h200monsta-glm-recipe-2026-10-10/summary.md",
     "quote": "| aligned-all | 190 | 0.31 | 0.34 | 0.47 | 166 | 444 | 419 | 0 | 221 | 1 | 2 | 0.72 | 302/442 |",
     "column": "Sam 1st tok, new turn"
    },
    "x_sam_turn_total_s": {
     "label": "Sam turn total, median",
     "value": 0.47,
     "unit": "s",
     "source": "fleet-bench/h200monsta-glm-recipe-2026-10-10/summary.md",
     "quote": "| aligned-all | 190 | 0.31 | 0.34 | 0.47 | 166 | 444 | 419 | 0 | 221 | 1 | 2 | 0.72 | 302/442 |",
     "column": "Sam turn total med"
    },
    "x_agg_decode_8stream": {
     "label": "Aggregate decode at 8 streams",
     "value": 444.0,
     "unit": "tok/s",
     "source": "fleet-bench/h200monsta-glm-recipe-2026-10-10/summary.md",
     "quote": "| aligned-all | 190 | 0.31 | 0.34 | 0.47 | 166 | 444 | 419 | 0 | 221 | 1 | 2 | 0.72 | 302/442 |",
     "column": "8-stream tok/s"
    },
    "x_agg_decode_16stream": {
     "label": "Aggregate decode at 16 streams",
     "value": 419.0,
     "unit": "tok/s",
     "source": "fleet-bench/h200monsta-glm-recipe-2026-10-10/summary.md",
     "quote": "| aligned-all | 190 | 0.31 | 0.34 | 0.47 | 166 | 444 | 419 | 0 | 221 | 1 | 2 | 0.72 | 302/442 |",
     "column": "16-stream tok/s"
    },
    "x_corpus_score": {
     "label": "GLM corpus score of 264",
     "value": 221,
     "unit": "count",
     "source": "fleet-bench/h200monsta-glm-recipe-2026-10-10/summary.md",
     "quote": "| aligned-all | 190 | 0.31 | 0.34 | 0.47 | 166 | 444 | 419 | 0 | 221 | 1 | 2 | 0.72 | 302/442 |",
     "column": "corpus /264"
    }
   },
   "notes": "TP2 recipe sweep config 'aligned-all' on h200monsta GPUs 1,3 (:9420). Decodes are single-stream Sam-turn measurements."
  },
  {
   "id": "rtxpro-2x-glm-5-3-flash-fp8-tp2-aligned-chunk1024",
   "model": "GLM-5.3-Flash",
   "model_slug": "glm-5-3-flash",
   "hardware": "2x RTX PRO 6000 Blackwell Max-Q (h200monsta)",
   "hardware_family": "rtx-pro-6000",
   "units": 2,
   "engine": "vLLM",
   "quant": "FP8",
   "measured": "2026-10-10",
   "metrics": {
    "decode_1u_prose": {
     "value": 159.0,
     "unit": "tok/s",
     "source": "fleet-bench/h200monsta-glm-recipe-2026-10-10/summary.md",
     "quote": "| aligned-chunk1024 | 221 | 0.55 | 0.58 | 0.68 | 159 | 379 | 402 | 0 | 221 | 1 | 2 | 1.01 | 295/300 |",
     "column": "decode tok/s"
    },
    "x_sam_first_token_s": {
     "label": "Sam first-token latency, new turn",
     "value": 0.58,
     "unit": "s",
     "source": "fleet-bench/h200monsta-glm-recipe-2026-10-10/summary.md",
     "quote": "| aligned-chunk1024 | 221 | 0.55 | 0.58 | 0.68 | 159 | 379 | 402 | 0 | 221 | 1 | 2 | 1.01 | 295/300 |",
     "column": "Sam 1st tok, new turn"
    },
    "x_sam_turn_total_s": {
     "label": "Sam turn total, median",
     "value": 0.68,
     "unit": "s",
     "source": "fleet-bench/h200monsta-glm-recipe-2026-10-10/summary.md",
     "quote": "| aligned-chunk1024 | 221 | 0.55 | 0.58 | 0.68 | 159 | 379 | 402 | 0 | 221 | 1 | 2 | 1.01 | 295/300 |",
     "column": "Sam turn total med"
    },
    "x_agg_decode_8stream": {
     "label": "Aggregate decode at 8 streams",
     "value": 379.0,
     "unit": "tok/s",
     "source": "fleet-bench/h200monsta-glm-recipe-2026-10-10/summary.md",
     "quote": "| aligned-chunk1024 | 221 | 0.55 | 0.58 | 0.68 | 159 | 379 | 402 | 0 | 221 | 1 | 2 | 1.01 | 295/300 |",
     "column": "8-stream tok/s"
    },
    "x_agg_decode_16stream": {
     "label": "Aggregate decode at 16 streams",
     "value": 402.0,
     "unit": "tok/s",
     "source": "fleet-bench/h200monsta-glm-recipe-2026-10-10/summary.md",
     "quote": "| aligned-chunk1024 | 221 | 0.55 | 0.58 | 0.68 | 159 | 379 | 402 | 0 | 221 | 1 | 2 | 1.01 | 295/300 |",
     "column": "16-stream tok/s"
    },
    "x_corpus_score": {
     "label": "GLM corpus score of 264",
     "value": 221,
     "unit": "count",
     "source": "fleet-bench/h200monsta-glm-recipe-2026-10-10/summary.md",
     "quote": "| aligned-chunk1024 | 221 | 0.55 | 0.58 | 0.68 | 159 | 379 | 402 | 0 | 221 | 1 | 2 | 1.01 | 295/300 |",
     "column": "corpus /264"
    }
   },
   "notes": "TP2 recipe sweep config 'aligned-chunk1024' on h200monsta GPUs 1,3 (:9420). Decodes are single-stream Sam-turn measurements."
  },
  {
   "id": "rtxpro-2x-glm-5-3-flash-fp8-tp2-aligned-chunk2048",
   "model": "GLM-5.3-Flash",
   "model_slug": "glm-5-3-flash",
   "hardware": "2x RTX PRO 6000 Blackwell Max-Q (h200monsta)",
   "hardware_family": "rtx-pro-6000",
   "units": 2,
   "engine": "vLLM",
   "quant": "FP8",
   "measured": "2026-10-10",
   "metrics": {
    "decode_1u_prose": {
     "value": 156.0,
     "unit": "tok/s",
     "source": "fleet-bench/h200monsta-glm-recipe-2026-10-10/summary.md",
     "quote": "| aligned-chunk2048 | 221 | 0.43 | 0.46 | 0.59 | 156 | 410 | 399 | 0 | 227 | 0 | 2 | 0.84 | 294/305 |",
     "column": "decode tok/s"
    },
    "x_sam_first_token_s": {
     "label": "Sam first-token latency, new turn",
     "value": 0.46,
     "unit": "s",
     "source": "fleet-bench/h200monsta-glm-recipe-2026-10-10/summary.md",
     "quote": "| aligned-chunk2048 | 221 | 0.43 | 0.46 | 0.59 | 156 | 410 | 399 | 0 | 227 | 0 | 2 | 0.84 | 294/305 |",
     "column": "Sam 1st tok, new turn"
    },
    "x_sam_turn_total_s": {
     "label": "Sam turn total, median",
     "value": 0.59,
     "unit": "s",
     "source": "fleet-bench/h200monsta-glm-recipe-2026-10-10/summary.md",
     "quote": "| aligned-chunk2048 | 221 | 0.43 | 0.46 | 0.59 | 156 | 410 | 399 | 0 | 227 | 0 | 2 | 0.84 | 294/305 |",
     "column": "Sam turn total med"
    },
    "x_agg_decode_8stream": {
     "label": "Aggregate decode at 8 streams",
     "value": 410.0,
     "unit": "tok/s",
     "source": "fleet-bench/h200monsta-glm-recipe-2026-10-10/summary.md",
     "quote": "| aligned-chunk2048 | 221 | 0.43 | 0.46 | 0.59 | 156 | 410 | 399 | 0 | 227 | 0 | 2 | 0.84 | 294/305 |",
     "column": "8-stream tok/s"
    },
    "x_agg_decode_16stream": {
     "label": "Aggregate decode at 16 streams",
     "value": 399.0,
     "unit": "tok/s",
     "source": "fleet-bench/h200monsta-glm-recipe-2026-10-10/summary.md",
     "quote": "| aligned-chunk2048 | 221 | 0.43 | 0.46 | 0.59 | 156 | 410 | 399 | 0 | 227 | 0 | 2 | 0.84 | 294/305 |",
     "column": "16-stream tok/s"
    },
    "x_corpus_score": {
     "label": "GLM corpus score of 264",
     "value": 227,
     "unit": "count",
     "source": "fleet-bench/h200monsta-glm-recipe-2026-10-10/summary.md",
     "quote": "| aligned-chunk2048 | 221 | 0.43 | 0.46 | 0.59 | 156 | 410 | 399 | 0 | 227 | 0 | 2 | 0.84 | 294/305 |",
     "column": "corpus /264"
    }
   },
   "notes": "TP2 recipe sweep config 'aligned-chunk2048' on h200monsta GPUs 1,3 (:9420). Decodes are single-stream Sam-turn measurements."
  },
  {
   "id": "rtxpro-2x-glm-5-3-flash-fp8-tp2-aligned-low",
   "model": "GLM-5.3-Flash",
   "model_slug": "glm-5-3-flash",
   "hardware": "2x RTX PRO 6000 Blackwell Max-Q (h200monsta)",
   "hardware_family": "rtx-pro-6000",
   "units": 2,
   "engine": "vLLM",
   "quant": "FP8",
   "measured": "2026-10-10",
   "metrics": {
    "decode_1u_prose": {
     "value": 152.0,
     "unit": "tok/s",
     "source": "fleet-bench/h200monsta-glm-recipe-2026-10-10/summary.md",
     "quote": "| aligned-low | 190 | 0.39 | 0.39 | 0.47 | 152 | 406 | 395 | 0 | 215 | 0 | 4 | 0.67 | 293/300 |",
     "column": "decode tok/s"
    },
    "x_sam_first_token_s": {
     "label": "Sam first-token latency, new turn",
     "value": 0.39,
     "unit": "s",
     "source": "fleet-bench/h200monsta-glm-recipe-2026-10-10/summary.md",
     "quote": "| aligned-low | 190 | 0.39 | 0.39 | 0.47 | 152 | 406 | 395 | 0 | 215 | 0 | 4 | 0.67 | 293/300 |",
     "column": "Sam 1st tok, new turn"
    },
    "x_sam_turn_total_s": {
     "label": "Sam turn total, median",
     "value": 0.47,
     "unit": "s",
     "source": "fleet-bench/h200monsta-glm-recipe-2026-10-10/summary.md",
     "quote": "| aligned-low | 190 | 0.39 | 0.39 | 0.47 | 152 | 406 | 395 | 0 | 215 | 0 | 4 | 0.67 | 293/300 |",
     "column": "Sam turn total med"
    },
    "x_agg_decode_8stream": {
     "label": "Aggregate decode at 8 streams",
     "value": 406.0,
     "unit": "tok/s",
     "source": "fleet-bench/h200monsta-glm-recipe-2026-10-10/summary.md",
     "quote": "| aligned-low | 190 | 0.39 | 0.39 | 0.47 | 152 | 406 | 395 | 0 | 215 | 0 | 4 | 0.67 | 293/300 |",
     "column": "8-stream tok/s"
    },
    "x_agg_decode_16stream": {
     "label": "Aggregate decode at 16 streams",
     "value": 395.0,
     "unit": "tok/s",
     "source": "fleet-bench/h200monsta-glm-recipe-2026-10-10/summary.md",
     "quote": "| aligned-low | 190 | 0.39 | 0.39 | 0.47 | 152 | 406 | 395 | 0 | 215 | 0 | 4 | 0.67 | 293/300 |",
     "column": "16-stream tok/s"
    },
    "x_corpus_score": {
     "label": "GLM corpus score of 264",
     "value": 215,
     "unit": "count",
     "source": "fleet-bench/h200monsta-glm-recipe-2026-10-10/summary.md",
     "quote": "| aligned-low | 190 | 0.39 | 0.39 | 0.47 | 152 | 406 | 395 | 0 | 215 | 0 | 4 | 0.67 | 293/300 |",
     "column": "corpus /264"
    }
   },
   "notes": "TP2 recipe sweep config 'aligned-low' on h200monsta GPUs 1,3 (:9420). Decodes are single-stream Sam-turn measurements."
  },
  {
   "id": "rtxpro-2x-glm-5-3-flash-fp8-tp2-aligned-mb256",
   "model": "GLM-5.3-Flash",
   "model_slug": "glm-5-3-flash",
   "hardware": "2x RTX PRO 6000 Blackwell Max-Q (h200monsta)",
   "hardware_family": "rtx-pro-6000",
   "units": 2,
   "engine": "vLLM",
   "quant": "FP8",
   "measured": "2026-10-10",
   "metrics": {
    "decode_1u_prose": {
     "value": 164.0,
     "unit": "tok/s",
     "source": "fleet-bench/h200monsta-glm-recipe-2026-10-10/summary.md",
     "quote": "| aligned-mb256 | 190 | 0.36 | 0.4 | 0.55 | 164 | 425 | 414 | 0 | 224 | 1 | 2 | 0.79 | 295/305 |",
     "column": "decode tok/s"
    },
    "x_sam_first_token_s": {
     "label": "Sam first-token latency, new turn",
     "value": 0.4,
     "unit": "s",
     "source": "fleet-bench/h200monsta-glm-recipe-2026-10-10/summary.md",
     "quote": "| aligned-mb256 | 190 | 0.36 | 0.4 | 0.55 | 164 | 425 | 414 | 0 | 224 | 1 | 2 | 0.79 | 295/305 |",
     "column": "Sam 1st tok, new turn"
    },
    "x_sam_turn_total_s": {
     "label": "Sam turn total, median",
     "value": 0.55,
     "unit": "s",
     "source": "fleet-bench/h200monsta-glm-recipe-2026-10-10/summary.md",
     "quote": "| aligned-mb256 | 190 | 0.36 | 0.4 | 0.55 | 164 | 425 | 414 | 0 | 224 | 1 | 2 | 0.79 | 295/305 |",
     "column": "Sam turn total med"
    },
    "x_agg_decode_8stream": {
     "label": "Aggregate decode at 8 streams",
     "value": 425.0,
     "unit": "tok/s",
     "source": "fleet-bench/h200monsta-glm-recipe-2026-10-10/summary.md",
     "quote": "| aligned-mb256 | 190 | 0.36 | 0.4 | 0.55 | 164 | 425 | 414 | 0 | 224 | 1 | 2 | 0.79 | 295/305 |",
     "column": "8-stream tok/s"
    },
    "x_agg_decode_16stream": {
     "label": "Aggregate decode at 16 streams",
     "value": 414.0,
     "unit": "tok/s",
     "source": "fleet-bench/h200monsta-glm-recipe-2026-10-10/summary.md",
     "quote": "| aligned-mb256 | 190 | 0.36 | 0.4 | 0.55 | 164 | 425 | 414 | 0 | 224 | 1 | 2 | 0.79 | 295/305 |",
     "column": "16-stream tok/s"
    },
    "x_corpus_score": {
     "label": "GLM corpus score of 264",
     "value": 224,
     "unit": "count",
     "source": "fleet-bench/h200monsta-glm-recipe-2026-10-10/summary.md",
     "quote": "| aligned-mb256 | 190 | 0.36 | 0.4 | 0.55 | 164 | 425 | 414 | 0 | 224 | 1 | 2 | 0.79 | 295/305 |",
     "column": "corpus /264"
    }
   },
   "notes": "TP2 recipe sweep config 'aligned-mb256' on h200monsta GPUs 1,3 (:9420). Decodes are single-stream Sam-turn measurements."
  },
  {
   "id": "rtxpro-2x-glm-5-3-flash-fp8-tp2-vram",
   "model": "GLM-5.3-Flash",
   "model_slug": "glm-5-3-flash",
   "hardware": "2x RTX PRO 6000 Blackwell Max-Q (h200monsta)",
   "hardware_family": "rtx-pro-6000",
   "units": 2,
   "engine": "vLLM",
   "quant": "FP8",
   "measured": "2026-10-10",
   "metrics": {
    "decode_1u_prose": {
     "value": 160.0,
     "unit": "tok/s",
     "source": "fleet-bench/h200monsta-glm-recipe-2026-10-10/summary.md",
     "quote": "| vram | 190 | 2.2 | 0.14 | 2.37 | 160 | 435 | 433 | 0 | 227 | 0 | 3 | 0.53 | 300/453 |",
     "column": "decode tok/s"
    },
    "x_sam_first_token_s": {
     "label": "Sam first-token latency, new turn",
     "value": 0.14,
     "unit": "s",
     "source": "fleet-bench/h200monsta-glm-recipe-2026-10-10/summary.md",
     "quote": "| vram | 190 | 2.2 | 0.14 | 2.37 | 160 | 435 | 433 | 0 | 227 | 0 | 3 | 0.53 | 300/453 |",
     "column": "Sam 1st tok, new turn"
    },
    "x_sam_turn_total_s": {
     "label": "Sam turn total, median",
     "value": 2.37,
     "unit": "s",
     "source": "fleet-bench/h200monsta-glm-recipe-2026-10-10/summary.md",
     "quote": "| vram | 190 | 2.2 | 0.14 | 2.37 | 160 | 435 | 433 | 0 | 227 | 0 | 3 | 0.53 | 300/453 |",
     "column": "Sam turn total med"
    },
    "x_agg_decode_8stream": {
     "label": "Aggregate decode at 8 streams",
     "value": 435.0,
     "unit": "tok/s",
     "source": "fleet-bench/h200monsta-glm-recipe-2026-10-10/summary.md",
     "quote": "| vram | 190 | 2.2 | 0.14 | 2.37 | 160 | 435 | 433 | 0 | 227 | 0 | 3 | 0.53 | 300/453 |",
     "column": "8-stream tok/s"
    },
    "x_agg_decode_16stream": {
     "label": "Aggregate decode at 16 streams",
     "value": 433.0,
     "unit": "tok/s",
     "source": "fleet-bench/h200monsta-glm-recipe-2026-10-10/summary.md",
     "quote": "| vram | 190 | 2.2 | 0.14 | 2.37 | 160 | 435 | 433 | 0 | 227 | 0 | 3 | 0.53 | 300/453 |",
     "column": "16-stream tok/s"
    },
    "x_corpus_score": {
     "label": "GLM corpus score of 264",
     "value": 227,
     "unit": "count",
     "source": "fleet-bench/h200monsta-glm-recipe-2026-10-10/summary.md",
     "quote": "| vram | 190 | 2.2 | 0.14 | 2.37 | 160 | 435 | 433 | 0 | 227 | 0 | 3 | 0.53 | 300/453 |",
     "column": "corpus /264"
    }
   },
   "notes": "TP2 recipe sweep config 'vram' on h200monsta GPUs 1,3 (:9420). Decodes are single-stream Sam-turn measurements."
  },
  {
   "id": "rtxpro-2x-glm-5-3-flash-fp8-tp2-vram-effort-low",
   "model": "GLM-5.3-Flash",
   "model_slug": "glm-5-3-flash",
   "hardware": "2x RTX PRO 6000 Blackwell Max-Q (h200monsta)",
   "hardware_family": "rtx-pro-6000",
   "units": 2,
   "engine": "vLLM",
   "quant": "FP8",
   "measured": "2026-10-10",
   "metrics": {
    "decode_1u_prose": {
     "value": 176.0,
     "unit": "tok/s",
     "source": "fleet-bench/h200monsta-glm-recipe-2026-10-10/summary.md",
     "quote": "| vram-effort-low | 200 | 2.2 | 0.14 | 2.32 | 176 | 434 | 454 | 0 | 221 | 0 | 2 | 0.41 | 301/457 |",
     "column": "decode tok/s"
    },
    "x_sam_first_token_s": {
     "label": "Sam first-token latency, new turn",
     "value": 0.14,
     "unit": "s",
     "source": "fleet-bench/h200monsta-glm-recipe-2026-10-10/summary.md",
     "quote": "| vram-effort-low | 200 | 2.2 | 0.14 | 2.32 | 176 | 434 | 454 | 0 | 221 | 0 | 2 | 0.41 | 301/457 |",
     "column": "Sam 1st tok, new turn"
    },
    "x_sam_turn_total_s": {
     "label": "Sam turn total, median",
     "value": 2.32,
     "unit": "s",
     "source": "fleet-bench/h200monsta-glm-recipe-2026-10-10/summary.md",
     "quote": "| vram-effort-low | 200 | 2.2 | 0.14 | 2.32 | 176 | 434 | 454 | 0 | 221 | 0 | 2 | 0.41 | 301/457 |",
     "column": "Sam turn total med"
    },
    "x_agg_decode_8stream": {
     "label": "Aggregate decode at 8 streams",
     "value": 434.0,
     "unit": "tok/s",
     "source": "fleet-bench/h200monsta-glm-recipe-2026-10-10/summary.md",
     "quote": "| vram-effort-low | 200 | 2.2 | 0.14 | 2.32 | 176 | 434 | 454 | 0 | 221 | 0 | 2 | 0.41 | 301/457 |",
     "column": "8-stream tok/s"
    },
    "x_agg_decode_16stream": {
     "label": "Aggregate decode at 16 streams",
     "value": 454.0,
     "unit": "tok/s",
     "source": "fleet-bench/h200monsta-glm-recipe-2026-10-10/summary.md",
     "quote": "| vram-effort-low | 200 | 2.2 | 0.14 | 2.32 | 176 | 434 | 454 | 0 | 221 | 0 | 2 | 0.41 | 301/457 |",
     "column": "16-stream tok/s"
    },
    "x_corpus_score": {
     "label": "GLM corpus score of 264",
     "value": 221,
     "unit": "count",
     "source": "fleet-bench/h200monsta-glm-recipe-2026-10-10/summary.md",
     "quote": "| vram-effort-low | 200 | 2.2 | 0.14 | 2.32 | 176 | 434 | 454 | 0 | 221 | 0 | 2 | 0.41 | 301/457 |",
     "column": "corpus /264"
    }
   },
   "notes": "TP2 recipe sweep config 'vram-effort-low' on h200monsta GPUs 1,3 (:9420). Decodes are single-stream Sam-turn measurements."
  },
  {
   "id": "rtxpro-4x-deepseek-v4-1-flash-nvfp4-tp4-ep",
   "model": "DeepSeek-V4.1-Flash (nvidia NVFP4)",
   "model_slug": "deepseek-v4-1-flash",
   "hardware": "4x RTX PRO 6000 Blackwell Max-Q (h200monsta)",
   "hardware_family": "rtx-pro-6000",
   "units": 4,
   "engine": "vLLM",
   "quant": "NVFP4",
   "measured": "2026-09-25",
   "metrics": {
    "decode_1u_prose": {
     "value": 61,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-25-deepseek-v41-results.md",
     "quote": "1 user **61 tok/s** (62 per stream, TTFT 0.05 s)"
    },
    "agg_decode_8u": {
     "value": 349,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-25-deepseek-v41-results.md",
     "quote": "8 users **349 aggregate** (46 per user, TTFT 0.29 s)"
    },
    "agg_decode_32u": {
     "value": 583,
     "unit": "tok/s",
     "source": "fleet-bench/2026-09-25-deepseek-v41-results.md",
     "quote": "32 users **583 aggregate** (38.5 per user, TTFT"
    },
    "context_verified": {
     "value": 163840,
     "unit": "tokens",
     "source": "fleet-bench/2026-09-25-deepseek-v41-results.md",
     "quote": "gmu 0.90, 163,840 context, 16 seqs"
    },
    "x_kv_cache_tokens": {
     "label": "KV cache capacity (fp8)",
     "value": 763467,
     "unit": "tokens",
     "source": "fleet-bench/2026-09-25-deepseek-v41-results.md",
     "quote": "MoE backend `FLASHINFER_CUTLASS` NvFp4. KV fp8 **763,467 tokens** (4.66 full 163k requests); 87.2 GB used per GPU while serving."
    },
    "toolcall_pass": {
     "value": 38,
     "unit": "count",
     "source": "fleet-bench/2026-09-25-deepseek-v41-results.md",
     "quote": "**Gate:** reasoning + `391`; tools 38/38."
    },
    "x_cmmc_closed_book_best": {
     "label": "CMMC closed-book score of 35 (best of 3 seeds; 35/33/33)",
     "value": 35,
     "unit": "count",
     "source": "fleet-bench/2026-09-25-deepseek-v41-results.md",
     "quote": "**CMMC 35, 33, 33 of 35**"
    }
   },
   "notes": "T9(b) second attempt on h200monsta (RTX PRO 6000 Blackwell x4): TP4 + expert parallelism, Engram on NVMe, 163,840 context, 16 seqs. Attempt 1 hung after the FlashInfer MoE autotune and is excluded."
  },
  {
   "id": "strix-halo-ai11-bonsai-2-27b-pq2",
   "model": "Bonsai 2 27B",
   "model_slug": "bonsai-2-27b",
   "hardware": "1x ai11 (Minisforum MS-S1 MAX, Ryzen AI Max+ PRO 495 / Radeon 8065S, 192 GB unified, 176 GiB GTT)",
   "hardware_family": "strix-halo",
   "units": 1,
   "engine": "PrismML llama.cpp fork b10772",
   "quant": "PQ2_0",
   "measured": "2026-10-13",
   "metrics": {
    "decode_1u_prose": {
     "value": 25,
     "unit": "tok/s",
     "source": "ai11/RESULTS-2026-10-13-model-sweep.md",
     "quote": "20/20, 25 tok/s"
    },
    "x_mmlu_pro": {
     "value": 32,
     "unit": "of 40",
     "label": "MMLU-Pro correct of 40 (60-item battery)",
     "source": "ai11/RESULTS-2026-10-13-model-sweep.md",
     "quote": "(32/40 + 20/20, 25 tok/s)"
    },
    "x_humaneval": {
     "value": 20,
     "unit": "of 20",
     "label": "HumanEval correct of 20 (60-item battery)",
     "source": "ai11/RESULTS-2026-10-13-model-sweep.md",
     "quote": "(32/40 + 20/20, 25 tok/s)"
    }
   }
  },
  {
   "id": "strix-halo-ai11-deepseek-v4-flash-0731-q8k-xl",
   "model": "DeepSeek-V4-Flash-0731",
   "model_slug": "deepseek-v4-flash-0731",
   "hardware": "1x ai11 (Minisforum MS-S1 MAX, Ryzen AI Max+ PRO 495 / Radeon 8065S, 192 GB unified, 176 GiB GTT)",
   "hardware_family": "strix-halo",
   "units": 1,
   "engine": "llama.cpp Vulkan",
   "quant": "Q8_K_XL",
   "measured": "2026-10-13",
   "metrics": {
    "decode_1u_prose": {
     "value": 10.5,
     "unit": "tok/s",
     "source": "ai11/RESULTS-2026-10-13-model-sweep.md",
     "quote": "DeepSeek-V4-Flash-0731 Q8_K_XL (10.5 tok/s; with its DSpark draft it was OOM-killed at\nload, so no speculative run fits)"
    }
   },
   "notes": "The DSpark speculative variant was OOM-killed at load, so only the base non-speculative run is recorded; no quality numbers were reported for it in the sweep."
  },
  {
   "id": "strix-halo-ai11-gemma-4-26b-a4b-q8",
   "model": "Gemma 4 26B-A4B",
   "model_slug": "gemma-4",
   "hardware": "1x ai11 (Minisforum MS-S1 MAX, Ryzen AI Max+ PRO 495 / Radeon 8065S, 192 GB unified, 176 GiB GTT)",
   "hardware_family": "strix-halo",
   "units": 1,
   "engine": "llama.cpp Vulkan",
   "quant": "Q8",
   "measured": "2026-10-13",
   "metrics": {
    "x_mmlu_pro": {
     "value": 14,
     "unit": "of 40",
     "label": "MMLU-Pro correct of 40 (60-item battery)",
     "source": "ai11/RESULTS-2026-10-13-model-sweep.md",
     "quote": "Gemma 4 26B-A4B (14/40,"
    }
   },
   "notes": "Provisional first pass at greedy temp 0: 28 of 60 answers hit the 16k cap (greedy looping); a vendor-sampling fairness rerun was queued."
  },
  {
   "id": "strix-halo-ai11-gemma-4-26b-a4b-qat-q4-mtp3",
   "model": "Gemma 4 26B-A4B QAT",
   "model_slug": "gemma-4",
   "hardware": "1x ai11 (Minisforum MS-S1 MAX, Ryzen AI Max+ PRO 495 / Radeon 8065S, 192 GB unified, 176 GiB GTT)",
   "hardware_family": "strix-halo",
   "units": 1,
   "engine": "llama.cpp Vulkan, MTP 3",
   "quant": "QAT Q4",
   "measured": "2026-10-13",
   "metrics": {
    "x_decode_low": {
     "value": 108,
     "unit": "tok/s",
     "label": "lowest of the measured generation range (fastest decoder in the sweep)",
     "source": "ai11/RESULTS-2026-10-13-model-sweep.md",
     "quote": "fastest decoder in the sweep at 108-118 tok/s"
    },
    "x_decode_high": {
     "value": 118,
     "unit": "tok/s",
     "label": "highest of the measured generation range (fastest decoder in the sweep)",
     "source": "ai11/RESULTS-2026-10-13-model-sweep.md",
     "quote": "fastest decoder in the sweep at 108-118 tok/s"
    },
    "x_mmlu_pro": {
     "value": 10,
     "unit": "of 40",
     "label": "MMLU-Pro correct of 40 (60-item battery)",
     "source": "ai11/RESULTS-2026-10-13-model-sweep.md",
     "quote": "but 10/40 with 32 of 60 answers at the 16k cap"
    }
   },
   "notes": "Fastest decoder in the sweep but quality collapsed to greedy looping at temp 0; vendor-sampling rerun queued."
  },
  {
   "id": "strix-halo-ai11-gemma-4-31b-dense-q8-mtp3",
   "model": "Gemma 4 31B dense",
   "model_slug": "gemma-4",
   "hardware": "1x ai11 (Minisforum MS-S1 MAX, Ryzen AI Max+ PRO 495 / Radeon 8065S, 192 GB unified, 176 GiB GTT)",
   "hardware_family": "strix-halo",
   "units": 1,
   "engine": "llama.cpp Vulkan, MTP 3",
   "quant": "Q8",
   "measured": "2026-10-13",
   "metrics": {
    "decode_1u_prose": {
     "value": 18,
     "unit": "tok/s",
     "source": "ai11/RESULTS-2026-10-13-model-sweep.md",
     "quote": "but 18 tok/s and 130 prefill"
    },
    "x_prefill": {
     "value": 130,
     "unit": "tok/s",
     "label": "prefill speed",
     "source": "ai11/RESULTS-2026-10-13-model-sweep.md",
     "quote": "but 18 tok/s and 130 prefill"
    },
    "x_mmlu_pro": {
     "value": 35,
     "unit": "of 40",
     "label": "MMLU-Pro correct of 40 (60-item battery)",
     "source": "ai11/RESULTS-2026-10-13-model-sweep.md",
     "quote": "dense Q8 + MTP 3 (35/40 + 19/20"
    },
    "x_humaneval": {
     "value": 19,
     "unit": "of 20",
     "label": "HumanEval correct of 20 (60-item battery)",
     "source": "ai11/RESULTS-2026-10-13-model-sweep.md",
     "quote": "dense Q8 + MTP 3 (35/40 + 19/20"
    }
   }
  },
  {
   "id": "strix-halo-ai11-glm-5-3-flash-iq4xs",
   "model": "GLM-5.3-Flash",
   "model_slug": "glm-5-3-flash",
   "hardware": "1x ai11 (Minisforum MS-S1 MAX, Ryzen AI Max+ PRO 495 / Radeon 8065S, 192 GB unified, 176 GiB GTT)",
   "hardware_family": "strix-halo",
   "units": 1,
   "engine": "llama.cpp Vulkan",
   "quant": "IQ4_XS",
   "measured": "2026-10-13",
   "metrics": {
    "decode_1u_prose": {
     "value": 15,
     "unit": "tok/s",
     "source": "ai11/RESULTS-2026-10-13-model-sweep.md",
     "quote": "but 15 tok/s, 173 prefill"
    },
    "x_prefill": {
     "value": 173,
     "unit": "tok/s",
     "label": "prefill speed",
     "source": "ai11/RESULTS-2026-10-13-model-sweep.md",
     "quote": "but 15 tok/s, 173 prefill"
    },
    "x_mmlu_pro": {
     "value": 36,
     "unit": "of 40",
     "label": "MMLU-Pro correct of 40 (60-item battery)",
     "source": "ai11/RESULTS-2026-10-13-model-sweep.md",
     "quote": "GLM-5.3-Flash IQ4_XS (36/40 + 19/20"
    },
    "x_humaneval": {
     "value": 19,
     "unit": "of 20",
     "label": "HumanEval correct of 20 (60-item battery)",
     "source": "ai11/RESULTS-2026-10-13-model-sweep.md",
     "quote": "GLM-5.3-Flash IQ4_XS (36/40 + 19/20"
    }
   },
   "notes": "Provisional first pass on one engine at greedy temp 0; a Gufo fairness rerun was queued."
  },
  {
   "id": "strix-halo-ai11-gpt-oss-120b-mxfp4",
   "model": "GPT-OSS 120B",
   "model_slug": "gpt-oss-120b",
   "hardware": "1x ai11 (Minisforum MS-S1 MAX, Ryzen AI Max+ PRO 495 / Radeon 8065S, 192 GB unified, 176 GiB GTT)",
   "hardware_family": "strix-halo",
   "units": 1,
   "engine": "llama.cpp Vulkan",
   "quant": "MXFP4",
   "measured": "2026-10-13",
   "metrics": {
    "decode_1u_prose": {
     "value": 41,
     "unit": "tok/s",
     "source": "ai11/RESULTS-2026-10-13-model-sweep.md",
     "quote": "| big model, default settings | GPT-OSS 120B MXFP4, Vulkan | 34/40 + 19/20, 41 tok/s; a reasoning-effort=high rerun is queued |",
     "column": "why",
     "part": 4
    },
    "x_mmlu_pro": {
     "value": 34,
     "unit": "of 40",
     "label": "MMLU-Pro correct of 40 (60-item battery, default reasoning effort)",
     "source": "ai11/RESULTS-2026-10-13-model-sweep.md",
     "quote": "| big model, default settings | GPT-OSS 120B MXFP4, Vulkan | 34/40 + 19/20, 41 tok/s; a reasoning-effort=high rerun is queued |",
     "column": "why",
     "part": 0
    },
    "x_humaneval": {
     "value": 19,
     "unit": "of 20",
     "label": "HumanEval correct of 20 (60-item battery, default reasoning effort)",
     "source": "ai11/RESULTS-2026-10-13-model-sweep.md",
     "quote": "| big model, default settings | GPT-OSS 120B MXFP4, Vulkan | 34/40 + 19/20, 41 tok/s; a reasoning-effort=high rerun is queued |",
     "column": "why",
     "part": 2
    }
   },
   "notes": "Default settings; a reasoning-effort=high rerun was queued in the fairness pass, so these are default-effort numbers."
  },
  {
   "id": "strix-halo-ai11-gpt-oss-20b-mxfp4",
   "model": "GPT-OSS 20B",
   "model_slug": "gpt-oss-20b",
   "hardware": "1x ai11 (Minisforum MS-S1 MAX, Ryzen AI Max+ PRO 495 / Radeon 8065S, 192 GB unified, 176 GiB GTT)",
   "hardware_family": "strix-halo",
   "units": 1,
   "engine": "llama.cpp Vulkan",
   "quant": "MXFP4",
   "measured": "2026-10-13",
   "metrics": {
    "decode_1u_prose": {
     "value": 55,
     "unit": "tok/s",
     "source": "ai11/RESULTS-2026-10-13-model-sweep.md",
     "quote": "| best small / fastest clean scorer | GPT-OSS 20B MXFP4, Vulkan (default reasoning effort) | 37/40 + 20/20, 0 truncated, 55 tok/s, 856 prefill, whole eval in 8.8 min (13 GB of weights) |",
     "column": "why",
     "part": 5
    },
    "x_prefill": {
     "value": 856,
     "unit": "tok/s",
     "label": "prefill speed",
     "source": "ai11/RESULTS-2026-10-13-model-sweep.md",
     "quote": "| best small / fastest clean scorer | GPT-OSS 20B MXFP4, Vulkan (default reasoning effort) | 37/40 + 20/20, 0 truncated, 55 tok/s, 856 prefill, whole eval in 8.8 min (13 GB of weights) |",
     "column": "why",
     "part": 6
    },
    "x_mmlu_pro": {
     "value": 37,
     "unit": "of 40",
     "label": "MMLU-Pro correct of 40 (60-item battery, default reasoning effort)",
     "source": "ai11/RESULTS-2026-10-13-model-sweep.md",
     "quote": "| best small / fastest clean scorer | GPT-OSS 20B MXFP4, Vulkan (default reasoning effort) | 37/40 + 20/20, 0 truncated, 55 tok/s, 856 prefill, whole eval in 8.8 min (13 GB of weights) |",
     "column": "why",
     "part": 0
    },
    "x_humaneval": {
     "value": 20,
     "unit": "of 20",
     "label": "HumanEval correct of 20 (60-item battery, default reasoning effort)",
     "source": "ai11/RESULTS-2026-10-13-model-sweep.md",
     "quote": "| best small / fastest clean scorer | GPT-OSS 20B MXFP4, Vulkan (default reasoning effort) | 37/40 + 20/20, 0 truncated, 55 tok/s, 856 prefill, whole eval in 8.8 min (13 GB of weights) |",
     "column": "why",
     "part": 3
    }
   }
  },
  {
   "id": "strix-halo-ai11-laguna-s-2-1-q4",
   "model": "Laguna-S-2.1",
   "model_slug": "laguna-s-2-1",
   "hardware": "1x ai11 (Minisforum MS-S1 MAX, Ryzen AI Max+ PRO 495 / Radeon 8065S, 192 GB unified, 176 GiB GTT)",
   "hardware_family": "strix-halo",
   "units": 1,
   "engine": "llama.cpp Vulkan",
   "quant": "Q4",
   "measured": "2026-10-13",
   "metrics": {
    "decode_1u_prose": {
     "value": 31,
     "unit": "tok/s",
     "source": "ai11/RESULTS-2026-10-13-model-sweep.md",
     "quote": "19/20, 31 tok/s, 134 prefill"
    },
    "x_prefill": {
     "value": 134,
     "unit": "tok/s",
     "label": "prefill speed",
     "source": "ai11/RESULTS-2026-10-13-model-sweep.md",
     "quote": "31 tok/s, 134 prefill"
    },
    "x_mmlu_pro": {
     "value": 32,
     "unit": "of 40",
     "label": "MMLU-Pro correct of 40 (60-item battery)",
     "source": "ai11/RESULTS-2026-10-13-model-sweep.md",
     "quote": "Laguna-S-2.1 Q4 (32/40 + 19/20"
    },
    "x_humaneval": {
     "value": 19,
     "unit": "of 20",
     "label": "HumanEval correct of 20 (60-item battery)",
     "source": "ai11/RESULTS-2026-10-13-model-sweep.md",
     "quote": "Laguna-S-2.1 Q4 (32/40 + 19/20"
    }
   },
   "notes": "Provisional first pass; a Gufo fairness rerun was queued in sweep14."
  },
  {
   "id": "strix-halo-ai11-mimo-v2-6-flash-mxfp4",
   "model": "MiMo-V2.6-Flash-RL",
   "model_slug": "mimo-v2-6-flash",
   "hardware": "1x ai11 (Minisforum MS-S1 MAX, Ryzen AI Max+ PRO 495 / Radeon 8065S, 192 GB unified, 176 GiB GTT)",
   "hardware_family": "strix-halo",
   "units": 1,
   "engine": "llama.cpp Vulkan",
   "quant": "MXFP4",
   "measured": "2026-10-13",
   "metrics": {
    "decode_1u_prose": {
     "value": 20,
     "unit": "tok/s",
     "source": "ai11/RESULTS-2026-10-13-model-sweep.md",
     "quote": "but 20 tok/s; with MTP it ran ai11 out"
    },
    "x_mmlu_pro": {
     "value": 36,
     "unit": "of 40",
     "label": "MMLU-Pro correct of 40 (60-item battery)",
     "source": "ai11/RESULTS-2026-10-13-model-sweep.md",
     "quote": "MXFP4 (36/40 + 20/20 but 20 tok/s"
    },
    "x_humaneval": {
     "value": 20,
     "unit": "of 20",
     "label": "HumanEval correct of 20 (60-item battery)",
     "source": "ai11/RESULTS-2026-10-13-model-sweep.md",
     "quote": "MXFP4 (36/40 + 20/20 but 20 tok/s"
    }
   },
   "notes": "The MTP variant ran the box out of RAM and froze it, so only the non-MTP run is recorded; a Gufo fairness rerun was queued."
  },
  {
   "id": "strix-halo-ai11-minicpm-v-4-7-35b-a3b-q8",
   "model": "MiniCPM-V 4.7 35B-A3B",
   "model_slug": "minicpm-v-4-7-35b-a3b",
   "hardware": "1x ai11 (Minisforum MS-S1 MAX, Ryzen AI Max+ PRO 495 / Radeon 8065S, 192 GB unified, 176 GiB GTT)",
   "hardware_family": "strix-halo",
   "units": 1,
   "engine": "llama.cpp Vulkan",
   "quant": "Q8_0",
   "measured": "2026-10-13",
   "metrics": {
    "decode_1u_prose": {
     "value": 58,
     "unit": "tok/s",
     "source": "ai11/RESULTS-2026-10-13-model-sweep.md",
     "quote": "| vision | MiniCPM-V 4.7 35B-A3B Q8_0, Vulkan | 58 tok/s, 35/40 + 19/20, 1.26k prefill (text-only tested here) |",
     "column": "why",
     "part": 0
    },
    "x_prefill": {
     "value": 1.26,
     "unit": "k tok/s",
     "label": "prefill speed",
     "source": "ai11/RESULTS-2026-10-13-model-sweep.md",
     "quote": "| vision | MiniCPM-V 4.7 35B-A3B Q8_0, Vulkan | 58 tok/s, 35/40 + 19/20, 1.26k prefill (text-only tested here) |",
     "column": "why",
     "part": 5
    },
    "x_mmlu_pro": {
     "value": 35,
     "unit": "of 40",
     "label": "MMLU-Pro correct of 40 (60-item battery, text-only)",
     "source": "ai11/RESULTS-2026-10-13-model-sweep.md",
     "quote": "| vision | MiniCPM-V 4.7 35B-A3B Q8_0, Vulkan | 58 tok/s, 35/40 + 19/20, 1.26k prefill (text-only tested here) |",
     "column": "why",
     "part": 1
    },
    "x_humaneval": {
     "value": 19,
     "unit": "of 20",
     "label": "HumanEval correct of 20 (60-item battery, text-only)",
     "source": "ai11/RESULTS-2026-10-13-model-sweep.md",
     "quote": "| vision | MiniCPM-V 4.7 35B-A3B Q8_0, Vulkan | 58 tok/s, 35/40 + 19/20, 1.26k prefill (text-only tested here) |",
     "column": "why",
     "part": 3
    }
   }
  },
  {
   "id": "strix-halo-ai11-nemotron-3-5-lightning-context",
   "model": "Nemotron-3.5-Lightning",
   "model_slug": "nemotron-3-5-lightning",
   "hardware": "1x ai11 (Minisforum MS-S1 MAX, Ryzen AI Max+ PRO 495 / Radeon 8065S, 192 GB unified, 176 GiB GTT)",
   "hardware_family": "strix-halo",
   "units": 1,
   "engine": "llama.cpp Vulkan",
   "quant": "UD-Q4_K_XL",
   "measured": "2026-10-16",
   "metrics": {
    "x_mmlu_pro": {
     "value": 37,
     "unit": "of 40",
     "label": "MMLU-Pro correct of 40 (60-item battery, best 4-bit picks class)",
     "source": "ai11/study-2026-10-16-fp8-fp16-use-cases.md",
     "quote": "UD-Q4_K_XL 37/40+18/20"
    }
   },
   "notes": "The 2026-10-16 study cites the sweep's best 4-bit scores (UD-Q4_K_XL 37/40+18/20). Per-model speed numbers for Nemotron on ai11 were not found in the source snapshot, so no speed metric is recorded here."
  },
  {
   "id": "strix-halo-ai11-ornith-1-0-35b-ud-q4-k-xl",
   "model": "Ornith-1.0-35B",
   "model_slug": "ornith-1-0-35b",
   "hardware": "1x ai11 (Minisforum MS-S1 MAX, Ryzen AI Max+ PRO 495 / Radeon 8065S, 192 GB unified, 176 GiB GTT)",
   "hardware_family": "strix-halo",
   "units": 1,
   "engine": "llama.cpp Vulkan",
   "quant": "UD-Q4_K_XL",
   "measured": "2026-10-13",
   "metrics": {
    "decode_1u_prose": {
     "value": 64,
     "unit": "tok/s",
     "source": "ai11/RESULTS-2026-10-13-model-sweep.md",
     "quote": "| tied for best score at speed | Ornith-1.0-35B UD-Q4_K_XL, Vulkan (no MTP draft published) | 37/40 + 18/20, 64 tok/s, 1.2k prefill |",
     "column": "why",
     "part": 4
    },
    "x_prefill": {
     "value": 1.2,
     "unit": "k tok/s",
     "label": "prefill speed",
     "source": "ai11/RESULTS-2026-10-13-model-sweep.md",
     "quote": "| tied for best score at speed | Ornith-1.0-35B UD-Q4_K_XL, Vulkan (no MTP draft published) | 37/40 + 18/20, 64 tok/s, 1.2k prefill |",
     "column": "why",
     "part": 5
    },
    "x_mmlu_pro": {
     "value": 37,
     "unit": "of 40",
     "label": "MMLU-Pro correct of 40 (60-item battery)",
     "source": "ai11/RESULTS-2026-10-13-model-sweep.md",
     "quote": "| tied for best score at speed | Ornith-1.0-35B UD-Q4_K_XL, Vulkan (no MTP draft published) | 37/40 + 18/20, 64 tok/s, 1.2k prefill |",
     "column": "why",
     "part": 0
    },
    "x_humaneval": {
     "value": 18,
     "unit": "of 20",
     "label": "HumanEval correct of 20 (60-item battery)",
     "source": "ai11/RESULTS-2026-10-13-model-sweep.md",
     "quote": "| tied for best score at speed | Ornith-1.0-35B UD-Q4_K_XL, Vulkan (no MTP draft published) | 37/40 + 18/20, 64 tok/s, 1.2k prefill |",
     "column": "why",
     "part": 2
    }
   }
  },
  {
   "id": "strix-halo-ai11-qwen3-6-35b-a3b-ud-q8k-xl-mtp3",
   "model": "Qwen3.6-35B-A3B",
   "model_slug": "qwen3-6-35b-a3b",
   "hardware": "1x ai11 (Minisforum MS-S1 MAX, Ryzen AI Max+ PRO 495 / Radeon 8065S, 192 GB unified, 176 GiB GTT)",
   "hardware_family": "strix-halo",
   "units": 1,
   "engine": "llama.cpp Vulkan, MTP 3",
   "quant": "UD-Q8_K_XL",
   "measured": "2026-10-13",
   "metrics": {
    "x_decode_low": {
     "value": 64,
     "unit": "tok/s",
     "label": "lowest of the measured generation range",
     "source": "ai11/RESULTS-2026-10-13-model-sweep.md",
     "quote": "| fastest good model | Qwen3.6-35B-A3B UD-Q8_K_XL, Vulkan, MTP 3 | 64-72 tok/s, 35/40 + 20/20, 1.07k prefill |",
     "column": "why",
     "part": 0
    },
    "x_decode_high": {
     "value": 72,
     "unit": "tok/s",
     "label": "highest of the measured generation range",
     "source": "ai11/RESULTS-2026-10-13-model-sweep.md",
     "quote": "| fastest good model | Qwen3.6-35B-A3B UD-Q8_K_XL, Vulkan, MTP 3 | 64-72 tok/s, 35/40 + 20/20, 1.07k prefill |",
     "column": "why",
     "part": 1
    },
    "x_prefill": {
     "value": 1.07,
     "unit": "k tok/s",
     "label": "prefill speed",
     "source": "ai11/RESULTS-2026-10-13-model-sweep.md",
     "quote": "| fastest good model | Qwen3.6-35B-A3B UD-Q8_K_XL, Vulkan, MTP 3 | 64-72 tok/s, 35/40 + 20/20, 1.07k prefill |",
     "column": "why",
     "part": 6
    },
    "x_mmlu_pro": {
     "value": 35,
     "unit": "of 40",
     "label": "MMLU-Pro correct of 40 (60-item battery)",
     "source": "ai11/RESULTS-2026-10-13-model-sweep.md",
     "quote": "| fastest good model | Qwen3.6-35B-A3B UD-Q8_K_XL, Vulkan, MTP 3 | 64-72 tok/s, 35/40 + 20/20, 1.07k prefill |",
     "column": "why",
     "part": 2
    },
    "x_humaneval": {
     "value": 20,
     "unit": "of 20",
     "label": "HumanEval correct of 20 (60-item battery)",
     "source": "ai11/RESULTS-2026-10-13-model-sweep.md",
     "quote": "| fastest good model | Qwen3.6-35B-A3B UD-Q8_K_XL, Vulkan, MTP 3 | 64-72 tok/s, 35/40 + 20/20, 1.07k prefill |",
     "column": "why",
     "part": 4
    }
   }
  },
  {
   "id": "strix-halo-ai11-qwen3-8-27b-fusion-q6k-mtp3",
   "model": "DavidAU Qwen3.8-27B fusion",
   "model_slug": "qwen3-8-27b",
   "hardware": "1x ai11 (Minisforum MS-S1 MAX, Ryzen AI Max+ PRO 495 / Radeon 8065S, 192 GB unified, 176 GiB GTT)",
   "hardware_family": "strix-halo",
   "units": 1,
   "engine": "llama.cpp (dense), MTP 3",
   "quant": "Q6_K",
   "measured": "2026-10-13",
   "metrics": {
    "x_mmlu_pro": {
     "value": 38,
     "unit": "of 40",
     "label": "MMLU-Pro correct of 40 (60-item battery)",
     "source": "ai11/RESULTS-2026-10-13-model-sweep.md",
     "quote": "| highest score, patient jobs | DavidAU Qwen3.8-27B fusion Q6_K, MTP 3 (dense) | 38/40 + 20/20 but 15-22 tok/s, 162 prefill |",
     "column": "why",
     "part": 0
    },
    "x_humaneval": {
     "value": 20,
     "unit": "of 20",
     "label": "HumanEval correct of 20 (60-item battery)",
     "source": "ai11/RESULTS-2026-10-13-model-sweep.md",
     "quote": "| highest score, patient jobs | DavidAU Qwen3.8-27B fusion Q6_K, MTP 3 (dense) | 38/40 + 20/20 but 15-22 tok/s, 162 prefill |",
     "column": "why",
     "part": 3
    },
    "x_decode_low": {
     "value": 15,
     "unit": "tok/s",
     "label": "lowest of the measured generation range",
     "source": "ai11/RESULTS-2026-10-13-model-sweep.md",
     "quote": "| highest score, patient jobs | DavidAU Qwen3.8-27B fusion Q6_K, MTP 3 (dense) | 38/40 + 20/20 but 15-22 tok/s, 162 prefill |",
     "column": "why",
     "part": 4
    },
    "x_decode_high": {
     "value": 22,
     "unit": "tok/s",
     "label": "highest of the measured generation range",
     "source": "ai11/RESULTS-2026-10-13-model-sweep.md",
     "quote": "| highest score, patient jobs | DavidAU Qwen3.8-27B fusion Q6_K, MTP 3 (dense) | 38/40 + 20/20 but 15-22 tok/s, 162 prefill |",
     "column": "why",
     "part": 5
    },
    "x_prefill": {
     "value": 162,
     "unit": "tok/s",
     "label": "prefill speed",
     "source": "ai11/RESULTS-2026-10-13-model-sweep.md",
     "quote": "| highest score, patient jobs | DavidAU Qwen3.8-27B fusion Q6_K, MTP 3 (dense) | 38/40 + 20/20 but 15-22 tok/s, 162 prefill |",
     "column": "why",
     "part": 6
    }
   },
   "notes": "Highest score in the sweep but slow: a patient-jobs dense model."
  },
  {
   "id": "strix-halo-ai11-qwen3-8-flash-next-gufo-mtp7-ud-q4-k-xl",
   "model": "Qwen3.8-Flash-Next",
   "model_slug": "qwen3-8-flash-next",
   "hardware": "1x ai11 (Minisforum MS-S1 MAX, Ryzen AI Max+ PRO 495 / Radeon 8065S, 192 GB unified, 176 GiB GTT)",
   "hardware_family": "strix-halo",
   "units": 1,
   "engine": "Gufo 0.12.0 (ROCm)",
   "quant": "UD-Q4_K_XL",
   "measured": "2026-10-12",
   "metrics": {
    "decode_1u_prose": {
     "value": 46.5,
     "unit": "tok/s",
     "source": "ai11/RESULTS-2026-10-12-qwen38-flash.md",
     "quote": "**Linux Gufo ROCm, MTP 7** | **46.5** | **47.9** | 39.3 | **1717**",
     "column": "chat"
    },
    "decode_1u_code": {
     "value": 47.9,
     "unit": "tok/s",
     "source": "ai11/RESULTS-2026-10-12-qwen38-flash.md",
     "quote": "**Linux Gufo ROCm, MTP 7** | **46.5** | **47.9** | 39.3 | **1717**",
     "column": "code 512"
    },
    "x_needle_gen_39k": {
     "value": 39.3,
     "unit": "tok/s",
     "label": "39k-token needle generation speed (MAGNOLIA passed)",
     "source": "ai11/RESULTS-2026-10-12-qwen38-flash.md",
     "quote": "**Linux Gufo ROCm, MTP 7** | **46.5** | **47.9** | 39.3 | **1717**",
     "column": "39k needle gen"
    },
    "x_prefill_39k": {
     "value": 1717,
     "unit": "tok/s",
     "label": "39k-token prompt prefill speed",
     "source": "ai11/RESULTS-2026-10-12-qwen38-flash.md",
     "quote": "**Linux Gufo ROCm, MTP 7** | **46.5** | **47.9** | 39.3 | **1717**",
     "column": "39k prefill"
    },
    "x_mmlu_pro": {
     "value": 37,
     "unit": "of 40",
     "label": "MMLU-Pro correct of 40 (60-item battery)",
     "source": "ai11/RESULTS-2026-10-13-model-sweep.md",
     "quote": "| best overall | Qwen3.8-Flash-Next UD-Q4_K_XL, Gufo, MTP draft 7 | 37/40 + 18/20, 46-48 tok/s, 1.7k tok/s prefill (fastest long-prompt reader) |",
     "column": "why",
     "part": 0
    },
    "x_humaneval": {
     "value": 18,
     "unit": "of 20",
     "label": "HumanEval correct of 20 (60-item battery)",
     "source": "ai11/RESULTS-2026-10-13-model-sweep.md",
     "quote": "| best overall | Qwen3.8-Flash-Next UD-Q4_K_XL, Gufo, MTP draft 7 | 37/40 + 18/20, 46-48 tok/s, 1.7k tok/s prefill (fastest long-prompt reader) |",
     "column": "why",
     "part": 2
    }
   },
   "notes": "Best overall pick of the ai11 sweep; tool call OK and MAGNOLIA needle OK. MTP draft-max 7 on Gufo."
  },
  {
   "id": "strix-halo-ai11-qwen3-8-flash-next-gufo-no-mtp-ud-q4-k-xl",
   "model": "Qwen3.8-Flash-Next",
   "model_slug": "qwen3-8-flash-next",
   "hardware": "1x ai11 (Minisforum MS-S1 MAX, Ryzen AI Max+ PRO 495 / Radeon 8065S, 192 GB unified, 176 GiB GTT)",
   "hardware_family": "strix-halo",
   "units": 1,
   "engine": "Gufo 0.12.0 (ROCm), no MTP",
   "quant": "UD-Q4_K_XL",
   "measured": "2026-10-12",
   "metrics": {
    "decode_1u_prose": {
     "value": 29.5,
     "unit": "tok/s",
     "source": "ai11/RESULTS-2026-10-12-qwen38-flash.md",
     "quote": "| Linux Gufo ROCm, no MTP | 29.5 | 29.8 | 28.5 | **1779** |",
     "column": "chat"
    },
    "decode_1u_code": {
     "value": 29.8,
     "unit": "tok/s",
     "source": "ai11/RESULTS-2026-10-12-qwen38-flash.md",
     "quote": "| Linux Gufo ROCm, no MTP | 29.5 | 29.8 | 28.5 | **1779** |",
     "column": "code 512"
    },
    "x_needle_gen_39k": {
     "value": 28.5,
     "unit": "tok/s",
     "label": "39k-token needle generation speed",
     "source": "ai11/RESULTS-2026-10-12-qwen38-flash.md",
     "quote": "| Linux Gufo ROCm, no MTP | 29.5 | 29.8 | 28.5 | **1779** |",
     "column": "39k needle gen"
    },
    "x_prefill_39k": {
     "value": 1779,
     "unit": "tok/s",
     "label": "39k-token prompt prefill speed",
     "source": "ai11/RESULTS-2026-10-12-qwen38-flash.md",
     "quote": "| Linux Gufo ROCm, no MTP | 29.5 | 29.8 | 28.5 | **1779** |",
     "column": "39k prefill"
    }
   },
   "notes": "Fastest measured prefill on ai11 for this model; MTP raised generation but not prefill."
  },
  {
   "id": "strix-halo-ai11-qwen3-8-flash-next-rocm-mtp3-ud-q4-k-xl",
   "model": "Qwen3.8-Flash-Next",
   "model_slug": "qwen3-8-flash-next",
   "hardware": "1x ai11 (Minisforum MS-S1 MAX, Ryzen AI Max+ PRO 495 / Radeon 8065S, 192 GB unified, 176 GiB GTT)",
   "hardware_family": "strix-halo",
   "units": 1,
   "engine": "unsloth llama.cpp b11505-mix (ROCm), MTP 3",
   "quant": "UD-Q4_K_XL",
   "measured": "2026-10-12",
   "metrics": {
    "decode_1u_prose": {
     "value": 28.4,
     "unit": "tok/s",
     "source": "ai11/RESULTS-2026-10-12-qwen38-flash.md",
     "quote": "| Linux llama.cpp ROCm, MTP 3 | 28.4 | 40.2 | 32.8 | 429 |",
     "column": "chat"
    },
    "decode_1u_code": {
     "value": 40.2,
     "unit": "tok/s",
     "source": "ai11/RESULTS-2026-10-12-qwen38-flash.md",
     "quote": "| Linux llama.cpp ROCm, MTP 3 | 28.4 | 40.2 | 32.8 | 429 |",
     "column": "code 512"
    },
    "x_needle_gen_39k": {
     "value": 32.8,
     "unit": "tok/s",
     "label": "39k-token needle generation speed",
     "source": "ai11/RESULTS-2026-10-12-qwen38-flash.md",
     "quote": "| Linux llama.cpp ROCm, MTP 3 | 28.4 | 40.2 | 32.8 | 429 |",
     "column": "39k needle gen"
    },
    "x_prefill_39k": {
     "value": 429,
     "unit": "tok/s",
     "label": "39k-token prompt prefill speed",
     "source": "ai11/RESULTS-2026-10-12-qwen38-flash.md",
     "quote": "| Linux llama.cpp ROCm, MTP 3 | 28.4 | 40.2 | 32.8 | 429 |",
     "column": "39k prefill"
    },
    "x_llama_bench_pp512": {
     "value": 570.9,
     "unit": "tok/s",
     "label": "llama-bench pp512 (Linux, ROCm)",
     "source": "ai11/RESULTS-2026-10-12-qwen38-flash.md",
     "quote": "ROCm pp512 570.9 / pp4096 564.5 / tg128 22.9"
    },
    "x_llama_bench_pp4096": {
     "value": 564.5,
     "unit": "tok/s",
     "label": "llama-bench pp4096 (Linux, ROCm)",
     "source": "ai11/RESULTS-2026-10-12-qwen38-flash.md",
     "quote": "ROCm pp512 570.9 / pp4096 564.5 / tg128 22.9"
    },
    "x_llama_bench_tg128": {
     "value": 22.9,
     "unit": "tok/s",
     "label": "llama-bench tg128 (Linux, ROCm)",
     "source": "ai11/RESULTS-2026-10-12-qwen38-flash.md",
     "quote": "ROCm pp512 570.9 / pp4096 564.5 / tg128 22.9"
    }
   }
  },
  {
   "id": "strix-halo-ai11-qwen3-8-flash-next-vulkan-mtp3-ud-q4-k-xl",
   "model": "Qwen3.8-Flash-Next",
   "model_slug": "qwen3-8-flash-next",
   "hardware": "1x ai11 (Minisforum MS-S1 MAX, Ryzen AI Max+ PRO 495 / Radeon 8065S, 192 GB unified, 176 GiB GTT)",
   "hardware_family": "strix-halo",
   "units": 1,
   "engine": "unsloth llama.cpp b11505-mix (Vulkan RADV), MTP 3",
   "quant": "UD-Q4_K_XL",
   "measured": "2026-10-12",
   "metrics": {
    "decode_1u_prose": {
     "value": 36.6,
     "unit": "tok/s",
     "source": "ai11/RESULTS-2026-10-12-qwen38-flash.md",
     "quote": "| Linux llama.cpp Vulkan RADV, MTP 3 | 36.6 | 45.4 | 41.8 | 401 |",
     "column": "chat"
    },
    "decode_1u_code": {
     "value": 45.4,
     "unit": "tok/s",
     "source": "ai11/RESULTS-2026-10-12-qwen38-flash.md",
     "quote": "| Linux llama.cpp Vulkan RADV, MTP 3 | 36.6 | 45.4 | 41.8 | 401 |",
     "column": "code 512"
    },
    "x_needle_gen_39k": {
     "value": 41.8,
     "unit": "tok/s",
     "label": "39k-token needle generation speed",
     "source": "ai11/RESULTS-2026-10-12-qwen38-flash.md",
     "quote": "| Linux llama.cpp Vulkan RADV, MTP 3 | 36.6 | 45.4 | 41.8 | 401 |",
     "column": "39k needle gen"
    },
    "x_prefill_39k": {
     "value": 401,
     "unit": "tok/s",
     "label": "39k-token prompt prefill speed",
     "source": "ai11/RESULTS-2026-10-12-qwen38-flash.md",
     "quote": "| Linux llama.cpp Vulkan RADV, MTP 3 | 36.6 | 45.4 | 41.8 | 401 |",
     "column": "39k prefill"
    }
   }
  },
  {
   "id": "strix-halo-ai11-qwen3-8-flash-next-vulkan-no-mtp-ud-q4-k-xl",
   "model": "Qwen3.8-Flash-Next",
   "model_slug": "qwen3-8-flash-next",
   "hardware": "1x ai11 (Minisforum MS-S1 MAX, Ryzen AI Max+ PRO 495 / Radeon 8065S, 192 GB unified, 176 GiB GTT)",
   "hardware_family": "strix-halo",
   "units": 1,
   "engine": "unsloth llama.cpp b11505-mix (Vulkan RADV), no MTP",
   "quant": "UD-Q4_K_XL",
   "measured": "2026-10-12",
   "metrics": {
    "decode_1u_prose": {
     "value": 27.1,
     "unit": "tok/s",
     "source": "ai11/RESULTS-2026-10-12-qwen38-flash.md",
     "quote": "| Linux llama.cpp Vulkan RADV, no MTP | 27.1 | 27.1 | 24.9 | 435 |",
     "column": "chat"
    },
    "decode_1u_code": {
     "value": 27.1,
     "unit": "tok/s",
     "source": "ai11/RESULTS-2026-10-12-qwen38-flash.md",
     "quote": "| Linux llama.cpp Vulkan RADV, no MTP | 27.1 | 27.1 | 24.9 | 435 |",
     "column": "code 512"
    },
    "x_needle_gen_39k": {
     "value": 24.9,
     "unit": "tok/s",
     "label": "39k-token needle generation speed",
     "source": "ai11/RESULTS-2026-10-12-qwen38-flash.md",
     "quote": "| Linux llama.cpp Vulkan RADV, no MTP | 27.1 | 27.1 | 24.9 | 435 |",
     "column": "39k needle gen"
    },
    "x_prefill_39k": {
     "value": 435,
     "unit": "tok/s",
     "label": "39k-token prompt prefill speed",
     "source": "ai11/RESULTS-2026-10-12-qwen38-flash.md",
     "quote": "| Linux llama.cpp Vulkan RADV, no MTP | 27.1 | 27.1 | 24.9 | 435 |",
     "column": "39k prefill"
    },
    "x_llama_bench_pp512": {
     "value": 577.6,
     "unit": "tok/s",
     "label": "llama-bench pp512 (Linux, Vulkan)",
     "source": "ai11/RESULTS-2026-10-12-qwen38-flash.md",
     "quote": "llama-bench: Vulkan pp512 577.6"
    },
    "x_llama_bench_pp4096": {
     "value": 565.7,
     "unit": "tok/s",
     "label": "llama-bench pp4096 (Linux, Vulkan)",
     "source": "ai11/RESULTS-2026-10-12-qwen38-flash.md",
     "quote": "Vulkan pp512 577.6 / pp4096 565.7"
    },
    "x_llama_bench_tg128": {
     "value": 27.5,
     "unit": "tok/s",
     "label": "llama-bench tg128 (Linux, Vulkan)",
     "source": "ai11/RESULTS-2026-10-12-qwen38-flash.md",
     "quote": "Vulkan pp512 577.6 / pp4096 565.7 / tg128 27.5"
    }
   }
  },
  {
   "id": "strix-halo-ai11-win-qwen3-8-flash-next-vulkan-mtp3-ud-q4-k-xl",
   "model": "Qwen3.8-Flash-Next",
   "model_slug": "qwen3-8-flash-next",
   "hardware": "1x ai11 (Minisforum MS-S1 MAX, Ryzen AI Max+ PRO 495 / Radeon 8065S, 192 GB unified, Windows 11 Pro dual-boot)",
   "hardware_family": "strix-halo",
   "units": 1,
   "engine": "llama.cpp Vulkan (Windows), MTP 3",
   "quant": "UD-Q4_K_XL",
   "measured": "2026-10-12",
   "metrics": {
    "decode_1u_prose": {
     "value": 32.0,
     "unit": "tok/s",
     "source": "ai11/RESULTS-2026-10-12-qwen38-flash.md",
     "quote": "| Windows llama.cpp Vulkan, MTP 3 | 32.0 | 42.0 | 30.9 | 370 |",
     "column": "chat"
    },
    "decode_1u_code": {
     "value": 42.0,
     "unit": "tok/s",
     "source": "ai11/RESULTS-2026-10-12-qwen38-flash.md",
     "quote": "| Windows llama.cpp Vulkan, MTP 3 | 32.0 | 42.0 | 30.9 | 370 |",
     "column": "code 512"
    },
    "x_needle_gen_39k": {
     "value": 30.9,
     "unit": "tok/s",
     "label": "39k-token needle generation speed",
     "source": "ai11/RESULTS-2026-10-12-qwen38-flash.md",
     "quote": "| Windows llama.cpp Vulkan, MTP 3 | 32.0 | 42.0 | 30.9 | 370 |",
     "column": "39k needle gen"
    },
    "x_prefill_39k": {
     "value": 370,
     "unit": "tok/s",
     "label": "39k-token prompt prefill speed",
     "source": "ai11/RESULTS-2026-10-12-qwen38-flash.md",
     "quote": "| Windows llama.cpp Vulkan, MTP 3 | 32.0 | 42.0 | 30.9 | 370 |",
     "column": "39k prefill"
    }
   }
  },
  {
   "id": "strix-halo-ai11-win-qwen3-8-flash-next-vulkan-no-mtp-ud-q4-k-xl",
   "model": "Qwen3.8-Flash-Next",
   "model_slug": "qwen3-8-flash-next",
   "hardware": "1x ai11 (Minisforum MS-S1 MAX, Ryzen AI Max+ PRO 495 / Radeon 8065S, 192 GB unified, Windows 11 Pro dual-boot)",
   "hardware_family": "strix-halo",
   "units": 1,
   "engine": "llama.cpp Vulkan (Windows), no MTP",
   "quant": "UD-Q4_K_XL",
   "measured": "2026-10-12",
   "metrics": {
    "decode_1u_prose": {
     "value": 23.9,
     "unit": "tok/s",
     "source": "ai11/RESULTS-2026-10-12-qwen38-flash.md",
     "quote": "| Windows llama.cpp Vulkan, no MTP | 23.9 | 24.8 | 23.0 | 390 |",
     "column": "chat"
    },
    "decode_1u_code": {
     "value": 24.8,
     "unit": "tok/s",
     "source": "ai11/RESULTS-2026-10-12-qwen38-flash.md",
     "quote": "| Windows llama.cpp Vulkan, no MTP | 23.9 | 24.8 | 23.0 | 390 |",
     "column": "code 512"
    },
    "x_needle_gen_39k": {
     "value": 23.0,
     "unit": "tok/s",
     "label": "39k-token needle generation speed",
     "source": "ai11/RESULTS-2026-10-12-qwen38-flash.md",
     "quote": "| Windows llama.cpp Vulkan, no MTP | 23.9 | 24.8 | 23.0 | 390 |",
     "column": "39k needle gen"
    },
    "x_prefill_39k": {
     "value": 390,
     "unit": "tok/s",
     "label": "39k-token prompt prefill speed",
     "source": "ai11/RESULTS-2026-10-12-qwen38-flash.md",
     "quote": "| Windows llama.cpp Vulkan, no MTP | 23.9 | 24.8 | 23.0 | 390 |",
     "column": "39k prefill"
    },
    "x_llama_bench_pp512": {
     "value": 545.7,
     "unit": "tok/s",
     "label": "llama-bench pp512 (Windows baseline build)",
     "source": "ai11/RESULTS-2026-10-12-qwen38-flash.md",
     "quote": "| llama-bench pp512 / pp4096 | 545.7 / 557.2 tok/s | n/a |",
     "column": "baseline (ggml-org b11552)",
     "part": 0
    },
    "x_llama_bench_pp4096": {
     "value": 557.2,
     "unit": "tok/s",
     "label": "llama-bench pp4096 (Windows baseline build)",
     "source": "ai11/RESULTS-2026-10-12-qwen38-flash.md",
     "quote": "| llama-bench pp512 / pp4096 | 545.7 / 557.2 tok/s | n/a |",
     "column": "baseline (ggml-org b11552)",
     "part": 1
    },
    "x_llama_bench_tg128": {
     "value": 24.9,
     "unit": "tok/s",
     "label": "llama-bench tg128 (Windows baseline build)",
     "source": "ai11/RESULTS-2026-10-12-qwen38-flash.md",
     "quote": "| llama-bench tg128 | 24.9 tok/s | n/a |",
     "column": "baseline (ggml-org b11552)"
    }
   }
  },
  {
   "id": "strix-halo-g1a-gemma-4-26b-q8-ollama",
   "model": "Gemma 4 26B",
   "model_slug": "gemma-4",
   "hardware": "1x g1a (HP ZBook Ultra G1a, Ryzen AI Max+ PRO 395 / Radeon 8060S, 128 GB, Windows 11 Pro)",
   "hardware_family": "strix-halo",
   "units": 1,
   "engine": "Ollama 0.34.0 (ROCm)",
   "quant": "Q8_0",
   "measured": "2026-09-27",
   "metrics": {
    "decode_1u_prose": {
     "value": 51.8,
     "unit": "tok/s",
     "label": "median gen t/s across the four cases on g1a",
     "source": "fleet-bench/2026-09-27-strix-vs-gb10/README.md",
     "quote": "| gemma4:26b (MoE) | 66.1 | 51.8 | 2777 | 1190 |",
     "column": "g1a gen"
    },
    "x_prefill_long": {
     "value": 1190,
     "unit": "tok/s",
     "label": "long-prefill prompt t/s on g1a",
     "source": "fleet-bench/2026-09-27-strix-vs-gb10/README.md",
     "quote": "| gemma4:26b (MoE) | 66.1 | 51.8 | 2777 | 1190 |",
     "column": "g1a prefill (long)"
    }
   }
  },
  {
   "id": "strix-halo-g1a-gemma-4-31b-dense-ollama",
   "model": "Gemma 4 31B dense",
   "model_slug": "gemma-4",
   "hardware": "1x g1a (HP ZBook Ultra G1a, Ryzen AI Max+ PRO 395 / Radeon 8060S, 128 GB, Windows 11 Pro)",
   "hardware_family": "strix-halo",
   "units": 1,
   "engine": "Ollama 0.34.0 (ROCm)",
   "quant": "registry default quant",
   "measured": "2026-09-27",
   "metrics": {
    "decode_1u_prose": {
     "value": 8.8,
     "unit": "tok/s",
     "label": "median gen t/s across the four cases on g1a",
     "source": "fleet-bench/2026-09-27-strix-vs-gb10/README.md",
     "quote": "| gemma4:31b (dense) | 10.5 | 8.8 | 699 | 182 |",
     "column": "g1a gen"
    },
    "x_prefill_long": {
     "value": 182,
     "unit": "tok/s",
     "label": "long-prefill prompt t/s on g1a",
     "source": "fleet-bench/2026-09-27-strix-vs-gb10/README.md",
     "quote": "| gemma4:31b (dense) | 10.5 | 8.8 | 699 | 182 |",
     "column": "g1a prefill (long)"
    }
   }
  },
  {
   "id": "strix-halo-g1a-gpt-oss-120b-mxfp4-llamabench-rocm",
   "model": "GPT-OSS 120B",
   "model_slug": "gpt-oss-120b",
   "hardware": "1x g1a (HP ZBook Ultra G1a, Ryzen AI Max+ PRO 395 / Radeon 8060S, 128 GB, Windows 11 Pro)",
   "hardware_family": "strix-halo",
   "units": 1,
   "engine": "llama.cpp b11213 ROCm/HIP",
   "quant": "MXFP4",
   "measured": "2026-09-27",
   "metrics": {
    "x_pp512_d0": {
     "value": 685,
     "unit": "tok/s",
     "label": "llama-bench pp512 at d0",
     "source": "fleet-bench/2026-09-27-strix-vs-gb10/README.md",
     "quote": "| g1a ROCm (r3, idle) | 685 | 40.4 | 587 | 37.5 | 359 | 32.3 |",
     "column": "pp512 d0"
    },
    "x_pp512_8k": {
     "value": 587,
     "unit": "tok/s",
     "label": "llama-bench pp512 at 8K",
     "source": "fleet-bench/2026-09-27-strix-vs-gb10/README.md",
     "quote": "| g1a ROCm (r3, idle) | 685 | 40.4 | 587 | 37.5 | 359 | 32.3 |",
     "column": "pp512 8K"
    },
    "x_pp512_32k": {
     "value": 359,
     "unit": "tok/s",
     "label": "llama-bench pp512 at 32K",
     "source": "fleet-bench/2026-09-27-strix-vs-gb10/README.md",
     "quote": "| g1a ROCm (r3, idle) | 685 | 40.4 | 587 | 37.5 | 359 | 32.3 |",
     "column": "pp512 32K"
    },
    "x_tg128_8k": {
     "value": 37.5,
     "unit": "tok/s",
     "label": "llama-bench tg128 at 8K",
     "source": "fleet-bench/2026-09-27-strix-vs-gb10/README.md",
     "quote": "| g1a ROCm (r3, idle) | 685 | 40.4 | 587 | 37.5 | 359 | 32.3 |",
     "column": "tg128 8K"
    }
   }
  },
  {
   "id": "strix-halo-g1a-gpt-oss-120b-mxfp4-ollama",
   "model": "GPT-OSS 120B",
   "model_slug": "gpt-oss-120b",
   "hardware": "1x g1a (HP ZBook Ultra G1a, Ryzen AI Max+ PRO 395 / Radeon 8060S, 128 GB, Windows 11 Pro)",
   "hardware_family": "strix-halo",
   "units": 1,
   "engine": "Ollama 0.34.0 (ROCm)",
   "quant": "MXFP4",
   "measured": "2026-09-27",
   "metrics": {
    "x_gen_median": {
     "value": 40.4,
     "unit": "tok/s",
     "label": "median generation speed across the four bench cases (llama.cpp ROCm, tg128 d0)",
     "source": "fleet-bench/2026-09-27-strix-vs-gb10/README.md",
     "quote": "| g1a ROCm (r3, idle) | 685 | 40.4 | 587 | 37.5 | 359 | 32.3 |",
     "column": "tg128 d0"
    },
    "x_tg128_32k": {
     "value": 32.3,
     "unit": "tok/s",
     "label": "tg128 at 32K context (llama.cpp ROCm)",
     "source": "fleet-bench/2026-09-27-strix-vs-gb10/README.md",
     "quote": "| g1a ROCm (r3, idle) | 685 | 40.4 | 587 | 37.5 | 359 | 32.3 |",
     "column": "tg128 32K"
    },
    "x_tg128_d0": {
     "value": 44.7,
     "unit": "tok/s",
     "label": "tg128 at d0 (llama.cpp Vulkan)",
     "source": "fleet-bench/2026-09-27-strix-vs-gb10/README.md",
     "quote": "| g1a Vulkan (r2, idle) | 628 | 44.7 | 450 | 40.7 | 92 (samples 242, 18, 16) | 4.7 |",
     "column": "tg128 d0"
    }
   },
   "notes": "On Windows Vulkan gpt-oss-120b collapses at 32K (tg 4.7); ROCm holds 32.3, so ROCm is the usable backend for this model at long context."
  },
  {
   "id": "strix-halo-g1a-qwen2-5-72b-q4km-ollama",
   "model": "Qwen2.5 72B Instruct",
   "model_slug": "qwen2-5-72b-instruct",
   "hardware": "1x g1a (HP ZBook Ultra G1a, Ryzen AI Max+ PRO 395 / Radeon 8060S, 128 GB, Windows 11 Pro)",
   "hardware_family": "strix-halo",
   "units": 1,
   "engine": "Ollama 0.34.0 (ROCm)",
   "quant": "Q4_K_M",
   "measured": "2026-09-27",
   "metrics": {
    "decode_1u_prose": {
     "value": 4.0,
     "unit": "tok/s",
     "label": "median gen t/s across the four cases on g1a",
     "source": "fleet-bench/2026-09-27-strix-vs-gb10/README.md",
     "quote": "| qwen2.5:72b q4_K_M (dense) | 4.6 | 4.0 | 379 | 98 |",
     "column": "g1a gen"
    },
    "x_prefill_long": {
     "value": 98,
     "unit": "tok/s",
     "label": "long-prefill prompt t/s on g1a",
     "source": "fleet-bench/2026-09-27-strix-vs-gb10/README.md",
     "quote": "| qwen2.5:72b q4_K_M (dense) | 4.6 | 4.0 | 379 | 98 |",
     "column": "g1a prefill (long)"
    }
   }
  },
  {
   "id": "strix-halo-g1a-qwen3-6-27b-dense-ollama",
   "model": "Qwen3.6 27B (dense)",
   "model_slug": "qwen3-6-27b-dense",
   "hardware": "1x g1a (HP ZBook Ultra G1a, Ryzen AI Max+ PRO 395 / Radeon 8060S, 128 GB, Windows 11 Pro)",
   "hardware_family": "strix-halo",
   "units": 1,
   "engine": "Ollama 0.34.0 (ROCm)",
   "quant": "registry default quant",
   "measured": "2026-09-27",
   "metrics": {
    "decode_1u_prose": {
     "value": 11.4,
     "unit": "tok/s",
     "label": "median gen t/s across the four cases on g1a",
     "source": "fleet-bench/2026-09-27-strix-vs-gb10/README.md",
     "quote": "| qwen3.6:27b (dense) | 12.6 | 11.4 | 775 | 148 |",
     "column": "g1a gen"
    },
    "x_prefill_long": {
     "value": 148,
     "unit": "tok/s",
     "label": "long-prefill prompt t/s on g1a",
     "source": "fleet-bench/2026-09-27-strix-vs-gb10/README.md",
     "quote": "| qwen3.6:27b (dense) | 12.6 | 11.4 | 775 | 148 |",
     "column": "g1a prefill (long)"
    }
   }
  },
  {
   "id": "strix-halo-g1a-qwen3-6-35b-a3b-q4k-m-llamabench-rocm",
   "model": "Qwen3.6-35B-A3B",
   "model_slug": "qwen3-6-35b-a3b",
   "hardware": "1x g1a (HP ZBook Ultra G1a, Ryzen AI Max+ PRO 395 / Radeon 8060S, 128 GB, Windows 11 Pro)",
   "hardware_family": "strix-halo",
   "units": 1,
   "engine": "llama.cpp b11213 ROCm/HIP",
   "quant": "Q4_K_M",
   "measured": "2026-09-27",
   "metrics": {
    "x_pp512_d0": {
     "value": 1200,
     "unit": "tok/s",
     "label": "llama-bench pp512 at d0",
     "source": "fleet-bench/2026-09-27-strix-vs-gb10/README.md",
     "quote": "| g1a ROCm (r3, idle) | 1200 | 54.3 | 1049 | 50.0 | 760 | 44.7 |",
     "column": "pp512 d0"
    },
    "x_tg128_d0": {
     "value": 54.3,
     "unit": "tok/s",
     "label": "llama-bench tg128 at d0",
     "source": "fleet-bench/2026-09-27-strix-vs-gb10/README.md",
     "quote": "| g1a ROCm (r3, idle) | 1200 | 54.3 | 1049 | 50.0 | 760 | 44.7 |",
     "column": "tg128 d0"
    },
    "x_pp512_8k": {
     "value": 1049,
     "unit": "tok/s",
     "label": "llama-bench pp512 at 8K",
     "source": "fleet-bench/2026-09-27-strix-vs-gb10/README.md",
     "quote": "| g1a ROCm (r3, idle) | 1200 | 54.3 | 1049 | 50.0 | 760 | 44.7 |",
     "column": "pp512 8K"
    },
    "x_tg128_8k": {
     "value": 50.0,
     "unit": "tok/s",
     "label": "llama-bench tg128 at 8K",
     "source": "fleet-bench/2026-09-27-strix-vs-gb10/README.md",
     "quote": "| g1a ROCm (r3, idle) | 1200 | 54.3 | 1049 | 50.0 | 760 | 44.7 |",
     "column": "tg128 8K"
    },
    "x_pp512_32k": {
     "value": 760,
     "unit": "tok/s",
     "label": "llama-bench pp512 at 32K",
     "source": "fleet-bench/2026-09-27-strix-vs-gb10/README.md",
     "quote": "| g1a ROCm (r3, idle) | 1200 | 54.3 | 1049 | 50.0 | 760 | 44.7 |",
     "column": "pp512 32K"
    },
    "x_tg128_32k": {
     "value": 44.7,
     "unit": "tok/s",
     "label": "llama-bench tg128 at 32K",
     "source": "fleet-bench/2026-09-27-strix-vs-gb10/README.md",
     "quote": "| g1a ROCm (r3, idle) | 1200 | 54.3 | 1049 | 50.0 | 760 | 44.7 |",
     "column": "tg128 32K"
    }
   }
  },
  {
   "id": "strix-halo-g1a-qwen3-6-35b-a3b-q4k-m-llamabench-vulkan",
   "model": "Qwen3.6-35B-A3B",
   "model_slug": "qwen3-6-35b-a3b",
   "hardware": "1x g1a (HP ZBook Ultra G1a, Ryzen AI Max+ PRO 395 / Radeon 8060S, 128 GB, Windows 11 Pro)",
   "hardware_family": "strix-halo",
   "units": 1,
   "engine": "llama.cpp b11213 Vulkan",
   "quant": "Q4_K_M",
   "measured": "2026-09-27",
   "metrics": {
    "x_pp512_d0": {
     "value": 1313,
     "unit": "tok/s",
     "label": "llama-bench pp512 at d0",
     "source": "fleet-bench/2026-09-27-strix-vs-gb10/README.md",
     "quote": "| g1a Vulkan (r2, idle) | 1313 | 59.4 | 1028 | 56.4 | 638 | 49.4 |",
     "column": "pp512 d0"
    },
    "x_tg128_d0": {
     "value": 59.4,
     "unit": "tok/s",
     "label": "llama-bench tg128 at d0",
     "source": "fleet-bench/2026-09-27-strix-vs-gb10/README.md",
     "quote": "| g1a Vulkan (r2, idle) | 1313 | 59.4 | 1028 | 56.4 | 638 | 49.4 |",
     "column": "tg128 d0"
    },
    "x_pp512_8k": {
     "value": 1028,
     "unit": "tok/s",
     "label": "llama-bench pp512 at 8K",
     "source": "fleet-bench/2026-09-27-strix-vs-gb10/README.md",
     "quote": "| g1a Vulkan (r2, idle) | 1313 | 59.4 | 1028 | 56.4 | 638 | 49.4 |",
     "column": "pp512 8K"
    },
    "x_tg128_8k": {
     "value": 56.4,
     "unit": "tok/s",
     "label": "llama-bench tg128 at 8K",
     "source": "fleet-bench/2026-09-27-strix-vs-gb10/README.md",
     "quote": "| g1a Vulkan (r2, idle) | 1313 | 59.4 | 1028 | 56.4 | 638 | 49.4 |",
     "column": "tg128 8K"
    },
    "x_pp512_32k": {
     "value": 638,
     "unit": "tok/s",
     "label": "llama-bench pp512 at 32K",
     "source": "fleet-bench/2026-09-27-strix-vs-gb10/README.md",
     "quote": "| g1a Vulkan (r2, idle) | 1313 | 59.4 | 1028 | 56.4 | 638 | 49.4 |",
     "column": "pp512 32K"
    },
    "x_tg128_32k": {
     "value": 49.4,
     "unit": "tok/s",
     "label": "llama-bench tg128 at 32K",
     "source": "fleet-bench/2026-09-27-strix-vs-gb10/README.md",
     "quote": "| g1a Vulkan (r2, idle) | 1313 | 59.4 | 1028 | 56.4 | 638 | 49.4 |",
     "column": "tg128 32K"
    }
   }
  },
  {
   "id": "strix-halo-g1a-qwen3-6-35b-a3b-q4k-m-ollama",
   "model": "Qwen3.6-35B-A3B",
   "model_slug": "qwen3-6-35b-a3b",
   "hardware": "1x g1a (HP ZBook Ultra G1a, Ryzen AI Max+ PRO 395 / Radeon 8060S, 128 GB, Windows 11 Pro)",
   "hardware_family": "strix-halo",
   "units": 1,
   "engine": "Ollama 0.34.0 (ROCm)",
   "quant": "Q4_K_M",
   "measured": "2026-09-27",
   "metrics": {
    "decode_1u_prose": {
     "value": 54.4,
     "unit": "tok/s",
     "label": "median gen t/s across the four cases on g1a",
     "source": "fleet-bench/2026-09-27-strix-vs-gb10/README.md",
     "quote": "| qwen3.6:35b (MoE) | 74.5 | 54.4 | 2333 | 980 |",
     "column": "g1a gen"
    },
    "x_prefill_long": {
     "value": 980,
     "unit": "tok/s",
     "label": "long-prefill prompt t/s on g1a",
     "source": "fleet-bench/2026-09-27-strix-vs-gb10/README.md",
     "quote": "| qwen3.6:35b (MoE) | 74.5 | 54.4 | 2333 | 980 |",
     "column": "g1a prefill (long)"
    }
   },
   "notes": "Ollama April five-model set, same files and build as spark4 on the NVIDIA side."
  }
 ]
}