DeepSeek-V4.1-Flashbenchmark results

Measured speed and quality

Speed (single user)

Source: fleet-bench/2026-09-25-dgx1-best-model-eval.md, fleet-bench/ds41-v20-dgx1-2026-10-20/RESULTS.md, fleet-bench/t1023-mia-ds1017/SUMMARY.md, fleet-bench/t1018-ds-vs-glm/SUMMARY.md, fleet-bench/2026-10-02-best-model-per-tier-results.md, fleet-bench/2026-09-25-deepseek-v41-results.md

Aggregate throughput

Source: fleet-bench/2026-09-25-dgx1-best-model-eval.md, fleet-bench/ds41-v20-dgx1-2026-10-20/RESULTS.md, fleet-bench/2026-10-02-best-model-per-tier-results.md, fleet-bench/2026-09-25-deepseek-v41-results.md

Prefill and latency

Source: fleet-bench/t1023-mia-ds1017/SUMMARY.md, fleet-bench/t1018-ds-vs-glm/SUMMARY.md, fleet-bench/2026-10-02-best-model-per-tier-results.md

Context

Source: fleet-bench/ds41-v20-dgx1-2026-10-20/RESULTS.md, fleet-bench/2026-09-25-deepseek-v41-results.md

Quality

Source: fleet-bench/ds41-v20-dgx1-2026-10-20/RESULTS.md, fleet-bench/t1023-mia-ds1017/SUMMARY.md, fleet-bench/t1018-ds-vs-glm/SUMMARY.md, fleet-bench/2026-09-25-deepseek-v41-results.md

Power and efficiency

Source: fleet-bench/2026-10-02-best-model-per-tier-results.md

Other measurements

Source: fleet-bench/2026-09-25-dgx1-best-model-eval.md, fleet-bench/ds41-v20-dgx1-2026-10-20/RESULTS.md, fleet-bench/t1023-mia-ds1017/SUMMARY.md, fleet-bench/t1018-ds-vs-glm/SUMMARY.md, fleet-bench/2026-09-25-deepseek-v41-results.md

# DeepSeek-V4.1-Flash

What it is

DeepSeek-V4.1-Flash is a large mixture-of-experts model with vision input and a native long-context window, and it is the strongest quality-per-dollar option we have measured for grounded compliance and agent work on hardware you own. It ships in several checkpoints: an official release build (FP8 attention plus MXFP4 experts), an NVIDIA NVFP4 build, and community EXL3 low-bit packs around a low bits-per-weight EXL3 pack that let it fit on a pair of DGX Spark units. Across every hardware family we benchmarked, it passed our tool-calling gates, handled images, and served at least 160k tokens of verified context, with a one-million-token needle test passing on our DGX Station lane.

The model is not the fastest thing on every box. On the GB10 Spark class it needs community kits (TensorFold or the MiaAI recipe) and tensor parallelism across two or four units to be usable at all, and even then a single Spark pair trades speed for quality against GLM-5.3-Flash in some lanes. On the DGX Station GB300 it is the fastest text model we have measured, full stop, once the pin-hot-experts hook and our own RTX PRO six-thousand cold-expert sidecar are attached. That split is the story of this page: same model, wildly different results depending on the hardware and the serving recipe.

Where it runs best

Our measurements cover four hardware families, and the honest ranking is clear.

DGX Station GB300 (dgx1), one or two "units". This is the best home for DeepSeek-V4.1-Flash by a wide margin. With the official release checkpoint and a patched vLLM it served 70 tok/s to a single user and 276 tok/s aggregate at 32 users. With the Al-ENGR v20 pin-hot-experts recipe (keeping about 77 percent of each layer's experts resident in HBM) plus our own RTX PRO six-thousand cold-expert peer tier, the same box reached 180 tok/s single-user, over a thousand tok/s aggregate at 24 streams, and passed a needle test at 973,352 prompt tokens. The catch: this is Craig's production box, and the v20 lane is left running on it, so capacity is shared.

4x H200 (h200monsta). The NVIDIA NVFP4 build with tensor parallelism plus expert parallelism loads fully in VRAM and is the best H200 story: 61 tok/s single user, 349 aggregate at 8 users, 583 at 32, with a perfect 35-of-35 closed-book CMMC run in one seed. It failed its first serving attempt (the FlashInfer MoE autotune hung), so budget one retry. Context is verified at 163,840 tokens, not the full million.

4x GB10, TP4 (spark1 through spark4). The community DeepSeek-V4-Flash-0731 kit at NVFP4 with a 12-sequence profile is the best two-plus-Spark speed we measured for this model: 54.1 tok/s prose and 94.2 tok/s code single-user, 202 tok/s aggregate at the 12-user profile cap. It passed 7 of 8 gate categories with a thinking-on recheck. This is a good team-level option when you have four Sparks idle.

2x GB10, TP2. Two configurations are worth knowing. The older vLLM DeepSeek-V4-Flash-Vision-Exp recipe is fast for a pair (47.8 tok/s prose single-user, 117 aggregate at its 6-user cap) but loses on quality: it scored 75 of 93 correct on our grounded CMMC RAG suite against 91 for TensorFold. The TensorFold engine with an EXL3 low-bit pack is the two-Spark pick: 65.7 tok/s prose single-user, 125.6 aggregate at 4 streams, and quality equal to the four-unit lane. The newer MiaAI stock recipe on the same pair measured slightly slower (54.9 prose) with a touch more quantization loss on the CMMC RAG seed, so if you run two Sparks we recommend the TensorFold route.

Where it does not run: we never got the 8x GB10 pod or single-Spark configurations serving this model in the record window, and the stock loaders on the Station pushed host shared memory past Grace RAM repeatedly before the Engram-on-disk patch. Do not plan on running this model on one Spark.

Measured speed

All numbers below are from our fleet harness, median of repeated runs, and every value is traceable to the source files named on this page.

Single-user decode (prose). The GB300 with the v20 recipe and peer tier leads at 180 tok/s. The four-Spark TP4 kit follows at 54.1 prose (94.2 code). Two Sparks with TensorFold reach 65.7 prose; the newer MiaAI stock quant on the same pair measures 54.9 prose and 78.9 code. The H200 NVFP4 lane measures 61. The two-Spark Vision-Exp recipe measures 47.8 prose and 43.5 code. The GB300 release checkpoint lane (before the v20 work) measured 70.

Aggregate throughput at the concurrency ceiling. GB300 v20 plus peer tier: 1,054.1 tok/s at 24 streams (878.5 at 16). GB300 release checkpoint: 276 at 32 users. H200 NVFP4: 349 at 8 users and 583 at 32 (with 16 sequences configured, half the requests queue, so tail TTFT grows). Four Sparks TP4: 202.0 at 12 users. Two Sparks TensorFold: 125.6 at 4 streams. Two Sparks Vision-Exp: 117.0 at 6 users. Note how the Spark lanes cap hard at their profile's sequence count; they are single-team lanes, not department servers.

Prefill (long prompts). The Spark lanes hold remarkably flat prefill: 2,220 / 2,156 / 1,826 tok/s at 8k / 32k / 128k on four units, 2,045 / 2,066 / 2,043 on the MiaAI two-Spark lane, and 1,732 / 1,682 / 1,484 on the Vision-Exp pair. The GB300 v20 lane prefills far faster still (the source records a 131k prompt answered in about five seconds and a 1M-token needle in 84 seconds), but per-token prefill rates on that lane are not in this model's records, so we state them in words only.

Concurrency behavior worth knowing. On the GB300 the peer tier changed the shape of the curve: at 24 streams, TTFT p50 fell from queueing delays to under a second, and aggregate rose by nearly half. On the H200 the 32-user run still completed with zero failures. On two Sparks, cold prefill measured in the the low-to-mid seventeen hundreds to high eighteen hundreds range on the TensorFold lane.

Quality and accuracy

This is where DeepSeek-V4.1-Flash earns its place, and where the cheaper Spark recipes show their cost.

Grounded CMMC RAG (the number that counts for compliance work). The GB300 v20 lane scored accuracy 91, 90, 90 and strict 88, 89, 88 across three seeds, with all trap items caught on every seed. The two-Spark TensorFold lane scored 91 correct and 90 strict on seed 21, matching the GB300 on that seed, though it fabricated 3 answers versus GLM's 1 in the same lane and it is a single seed. The MiaAI stock quant scored 87 correct and 84 strict on the same seed, one notch below its TensorFold sibling. The older two-Spark Vision-Exp recipe collapsed to 75 correct and 70 strict with zero fabrication control, which is why we do not recommend it for compliance work.

Closed-book CMMC. The H200 NVFP4 lane posted 35, 33, and 33 of 35 across three seeds; the first 35 was the first perfect closed-book run in our entire benchmark program. The GB300 v20 lane scored 31 of 35 on the full battery (with repeat seeds at 30 and 31), and the GB300 release checkpoint scored 31, 33, 32. One honest caveat from the DGX Station gate: the model once invented a NIST 800-171 control requirement that does not exist instead of saying it does not exist. A grounded pipeline catches this; closed-book use should not trust it blindly.

Tool calling and agents. The H200 lane passed the full 38-case tools gate. The GB300 v20 lane passed 37 of 38, and on the harder 27-case multi-turn toolv2 suite it scored 24 of 27, tying the suite's best. The two-Spark lanes scored 20 of 27 on toolv2, below GLM where GLM was measured. On the grounded agent RAG suite, seeds other than 21 on the TensorFold lane were not run, so we cannot claim multi-seed parity there.

Coding. The GB300 v20 lane scored 10 of 10 on our code battery and 98.8 percent on HumanEval (vendor sampling). Four Sparks TP4 measured 94.2 tok/s code decode single-user and over three hundred fifty aggregate at 12 users, the strongest code throughput of any Spark lane we have on record for any model.

Vision. Both the H200 and the GB300 release checkpoint passed our vision smoke test 3 of 3. The two-Spark Vision-Exp recipe is vision-first by design. Image input is a genuine strength of this model across configurations.

How we ran it

Every number on this page comes from the same discipline: a correctness gate before any speed number is recorded, three repetitions with the median kept, launch flags recorded, and container logs kept before any container is removed. Speed ladders use 256-token outputs. Quality suites use vendor sampling at the engine's recommended settings unless noted.

The hardware configurations behind the records: one DGX Station GB300 (with an optional RTX PRO six-thousand Blackwell sidecar serving cold experts over shared memory), four H200s in tensor parallel plus expert parallel, four DGX Spark units in tensor parallel, and two DGX Spark units in tensor parallel under three different engines (vLLM with the Vision-Exp recipe, TensorFold with a low-bit EXL3 pack, and the MiaAI native engine with its stock experimental quant). The v20 GB300 lane keeps hot experts pinned in HBM and routes the rest to the peer GPU; that mechanism, not raw silicon, is why its numbers jumped 2.3x over our previous DGX Station lane.

Failures are part of the record: the H200 lane's first attempt hung during FlashInfer MoE autotune and is excluded; the DGX Station's stock loaders were killed six times by shared-memory exhaustion before the Engram-on-disk patch; the two-Spark MiaAI lane at maximum context ran host memory down to its guard floor and lost a rank until it was relaunched with reduced context and a self-healing service. The Spark pods never served the full NVFP4 checkpoint in the record window. We publish what worked and what did not, at the dates marked.

When to pick it

Pick DeepSeek-V4.1-Flash when grounded accuracy and long context matter more thanSeats-per-dollar, and you have at least an H200 x4 or a DGX Station to run it on. It is our best measured model for grounded compliance RAG, it passes the hardest tool-calling suite we run, it reads images, and it survived a real 1M-token needle. For a compliance or private-AI engagement where the model must cite retrieved documents correctly, it is the first model we reach for after the GLM family, and on the GB300 it is simply the fastest thing we have measured.

Pick something else when: you only have one or two DGX Sparks and need compliance-grade output (GLM-5.3-Flash on two Sparks matches its grounded accuracy there with simpler serving); you need a 32-seat department server on modest hardware (the model's Spark profiles cap at 6 or 12 users); or your workload is short-prompt chat where a smaller flash model on a single Spark is cheaper per token. Also think twice before mixing its quantizations: the community low-bit packs trade a measurable slice of grounded accuracy for fit, and the difference showed up in our own RAG seeds.

Where to go next

• LLM benchmarks hub
• Leaderboards
• How we measure
• Model tier list
• Private AI solutions
• Private LLM guide
• CMMC compliance

FAQ