Open-Weight LLM Tier ListThe Best Model for Every Hardware Budget We Measured
A tier answers one question: if you own this hardware (or are about to buy it), which open-weight model should you run for private work? The reason next to each placement is one line drawn from our own measured data, the same numbers shown on the model pages and the leaderboards.
These Tiers Are Our Judgment, Drawn From Measured Data
There is no formula that turns a benchmark table into a tier. Two teams looking at the same numbers weight them differently: a solo developer counts single-user tokens per second, a help desk counts aggregate throughput at thirty users, a compliance team counts retrieval accuracy against its own document set. Tier placement is therefore a call we made, and we say so plainly. What is not a judgment call is the data underneath: every speed, concurrency, context and accuracy number quoted here was measured by us on hardware we own, and each one can be traced to a stored result file. The rules for how we measure, including the correctness gate an engine must pass before we record any throughput, are on the methodology page.
What the tiers optimize for is private business use: a model that runs inside your own compliance boundary on hardware you own, serves the number of people who need it, answers from your own retrieved documents rather than inventing, and does not cost more than the problem is worth. That is why retrieval-grounded accuracy (what we call CMMC RAG accuracy in our battery) counts as heavily as raw speed, and why a configuration that cannot serve more than a handful of concurrent users is marked down for a team deployment even when its single-user speed is excellent.
Tiers mean: S = run this, it is the configuration we would deploy first at this budget. A = a strong choice with a real trade-off you should understand. B = good in a niche, wrong as a default. C = measured, and beaten by something cheaper or simpler for the same job. A model can land in different tiers on different hardware, and often does.
Want these numbers for your own hardware? Ask us to test a model on your hardware or call Penny at 919-348-4912.
Model Tiers by Hardware Budget
One section per hardware configuration we measured. Each entry names the model, its tier, and the one-line reason from the data. Links go to the full model page, where every number has its source file.
How to read the tiers (S is the best pick for that hardware, C the weakest):
One GB10 appliance: Gemma-4-26B-A4B (quality) and Nemotron-3.5-Lightning (throughput)
Why S: Gemma-4-26B-A4B is the quality pick on one box: 37 of 38 tool tasks and the best grounded compliance of the one-unit models at 62 percent strict on our CMMC RAG battery, with a verified context of 243,271 tokens, at 31.5 tok/s for one user and 866.4 tok/s across 64 concurrent users. Nemotron-3.5-Lightning is the speed pick: 90.2 tok/s prose for one user and 795.4 tok/s at its 64-user ceiling, at 6.3 tok/J, the most efficient layout we measured. Together they cover both jobs one appliance gets asked to do.
The catch: Nemotron's grounded accuracy is the weakest we measured, so it serves drafting and summarization, not compliance retrieval. Full data: the Gemma-4 model page and the Nemotron model page.
One GB10 appliance: gpt-oss-120b (accuracy-first, single user)
Why A: the best grounded compliance score of any one-unit model, 77 percent strict on the CMMC RAG battery, with a verified 120,755-token context, at 35.1 tok/s for one user.
The catch: it caps at 16 concurrent users and 198.0 tok/s aggregate at that ceiling, the lowest concurrency on the tier, and 1.63 tok/J is a quarter of Gemma's efficiency. Right for one careful analyst, wrong for a team.
One GB10 appliance: Qwen3.8-27B
Why B: a verified context of 243,392 tokens on one box, but 24.4 tok/s prose for a single user is the slowest of the one-unit NVFP4 models we measured.
The catch: Gemma matches the context, beats it on speed, concurrency and grounded accuracy, so Qwen3.8-27B only wins if your workflow is specifically tuned to it. Details on the Qwen3.8-27B model page.
Two GB10 units (pair): Qwen3.8-Flash-Next
Why S: the throughput pick for a pair: 329.6 tok/s across 32 concurrent users at a 10 tok/s floor, and the second-best grounded retrieval result in the whole fleet on the agentic CMMC RAG battery, with 52.6 tok/s code for one user.
The catch: a pair buys concurrency, not depth: 1.16 tok/J, and its context is configured at 131,072 tokens rather than a million. Full data on the Qwen3.8-Flash-Next model page.
Two GB10 units (pair): DeepSeek-V4-Flash-Vision-Exp
Why A: the best two-unit single-user model at 47.8 tok/s prose and the only pair that served a full million-token context, which makes it the long-document pick at this budget.
The catch: the serving profile caps at 6 concurrent sequences (117.0 tok/s aggregate at that ceiling), the least efficient layout measured at 0.41 tok/J, so it is a single-power-user box, not a team server.
Two GB10 units (pair): GLM-5.3-Flash
Why C: on two units GLM-5.3-Flash gives up too much for what it returns: 24.1 tok/s prose, 6 concurrent users, 68.3 tok/s at the ceiling and 0.23 tok/J, all the weakest pair numbers we measured.
The catch: none of this says anything bad about GLM-5.3-Flash itself; it says this model wants more memory bandwidth than a two-unit tensor-parallel pair provides. On four units or a GB300 it is a different story, as the rows below show.
Four-unit GB10 pod: GLM-5.3-Flash (quality) and Qwen3.8-Flash-Next (throughput)
Why S: the pod runs whichever the workload needs. GLM-5.3-Flash is the quality pick: 44.9 tok/s prose, a full million-token context, and the best closed-book compliance of any GB10 layout at 26 of 35 with zero hallucinated tool calls. Qwen3.8-Flash-Next is the throughput pick: 648.7 tok/s across 64 concurrent users and the pod's best prefill at 3,571 tok/s for 8k prompts.
The catch: the 200G RoCE fabric is not optional; the same four units on direct cables measured slower than one unit. Model detail: GLM-5.3-Flash model page.
Four-unit GB10 pod: DeepSeek-V4-Flash-0731
Why A: the pod's single-user speed leader at 54.1 tok/s prose and 94.2 tok/s code, with a full million-token context served on four units.
The catch: concurrency caps at 12 users and 202.0 tok/s aggregate at that ceiling, and 0.35 tok/J makes it the least efficient pod layout, so it earns its place only when one or two people need a big model fast.
DGX Station GB300: GLM-5.3-Flash (frontier work) and gpt-oss-120b (speed)
Why S: GLM-5.3-Flash on this box is the frontier pick: 145.9 tok/s for one user, a needle verified at a depth of 968,377 tokens (the deepest verified context we have measured anywhere in the fleet), 38 of 38 tool calls, 26 of 35 closed-book compliance and 2,517 tok/s at peak measured concurrency. gpt-oss-120b is the single-user speed king at 390.7 tok/s prose with 4,924.2 tok/s at 32 users, when the job is volume, not frontier quality.
The catch: GLM's verified-context crown comes from the nvidia NVFP4 checkpoint, which trades some exact compliance recall against the FP8 release run; and the box itself has no published MSRP, so budget on the reported launch figure. Full data on the GLM-5.3-Flash model page and the gpt-oss-120b model page.
DGX Station GB300: Qwen3.8-Flash-Next, Qwen3.6-35B-A3B and DeepSeek-V4.1-Flash (peer tier)
Why A: this is where the raw speed records live: Qwen3.6-35B-A3B in BF16 hits 335 tok/s for one user and 10,653 tok/s peak aggregate; Qwen3.8-Flash-Next hits 205 tok/s one-user and 7,040 peak aggregate; DeepSeek-V4.1-Flash in the peer-tier layout serves a verified context of 973,352 tokens with 37 of 38 tool calls at 180 tok/s.
The catch: closed-book compliance is the weak row for the fast ones (11 of 35 for Qwen3.8-Flash-Next, unrecorded for Qwen3.6-35B-A3B), so they serve coding and agentic pipelines, not compliance answers. DeepSeek's peer-tier layout also depends on a second card for cold experts.
DGX Station GB300: everything we measured but would not lead with
Why B: real measured results that lose their matchup on this box. Llama-3.3-70B BF16 manages 38.1 tok/s; Gemma-4-31B in BF16 manages 73.7 tok/s against 110.0 for its NVFP4 build; MiMo-V2.6-Flash-RL posts 212 tok/s but only 16 of 35 closed-book compliance; the full GLM-5.3 with experts offloaded to LPDDR5X drops to 34.5 tok/s.
The catch: each loses to a cheaper-or-faster row above for the same job. The exception worth naming: DeepSeek-V4.1-Flash in the FP8-plus-MXFP4 release layout posts the best closed-book compliance ever measured on this box at 31 of 35, but pays for it at 70 tok/s and 276 tok/s at 32 users.
Four H200s: GLM-5.3-Flash FP8 and Qwen3.8-27B FP8 for retrieval accuracy
Why S: GLM-5.3-Flash FP8 on the four-way NVLink pool is the overall concurrency and quality leader: 238 tok/s one user, 5,245 tok/s at peak measured concurrency, a verified context of 510,997 tokens, 90 percent strict on the CMMC RAG battery and zero measured hallucination rate. Qwen3.8-27B FP8 matches that 90 percent RAG strict score (the best RAG strict score we measured on the October 3 ladder) while serving 1,358.0 tok/s peak aggregate from a single card.
The catch: the pool is our production tier and priced accordingly; a business that fits inside a GB300's numbers should start there. Data on the DeepSeek-V4.1-Flash model page shows the H200 NVFP4 alternative at 583 tok/s across 32 users.
Four H200s: gpt-oss-120b
Why A: the throughput-per-dollar pick on this tier: 179 tok/s for one user, 1,511.2 tok/s peak across 64 users, and the fastest time to first token of the H200 cards we ran (0.904 s single user).
The catch: grounded accuracy is the weak row: 61 to 76 percent strict on the CMMC RAG battery against 90 for the GLM and Qwen FP8 cards, so it routes volume work, not compliance retrieval.
Four H200s: oversized single-card layouts
Why B: running one card of the pool for a small model is a waste of the other three. Qwen3.8-27B in the SEO-shaped 64-sequence layout peaks at 270.9 tok/s aggregate with 46.956 s time to first token at 16 users, numbers that lose to the same card correctly configured on the same box.
The catch: the card is fine; the configuration is not. This is our reminder that serving config moves results more than model swaps, which the methodology page documents in full.
Strix Halo (ai11 class): Qwen3.8-Flash-Next with the Gufo MTP7 recipe
Why S: the best measured recipe on the unified-memory box: 46.5 tok/s prose and 47.9 tok/s code for one user, with needle generation measured at 39.3 tok/s and 1,717 tok/s prefill at that depth. On a silent desktop with 176 GiB of shared memory, that is remarkable.
The catch: it is a single-user machine; there is no concurrent-serving story at this tier, and quality rows (MMLU-Pro 37, HumanEval 18) trail the big boxes. Recipe results live on the Qwen3.8-Flash-Next model page.
Strix Halo: gpt-oss-120b MXFP4 and the 35B A3B family
Why A: gpt-oss-120b runs whole in unified memory at 41 tok/s (ai11) and about 40 tok/s median generation on the g1a laptop, which makes it the biggest-model pick at this budget. Qwen3.6-35B-A3B in Q8 reaches 64 to 72 tok/s decode on ai11.
The catch: the Vulkan and ROCm prefill paths collapse to a few hundred tok/s on long prompts against the Gufo recipe's 1,717, so the recipe matters more than the model here.
Strix Halo: dense 27B builds and everything without a drafter
Why C: the dense fusion build of Qwen3.8-27B decodes at 15 to 22 tok/s with 162 tok/s prefill, and the same model without multi-token prediction drops to 27.1 tok/s. On this tier a missing drafter costs about a third of the speed.
The catch: these numbers are not wrong, just dominated: the same box with the right recipe runs the same class of work about 70 percent faster.
Where a model has no measured value for a field we say so rather than estimate. Configurations we measured but did not tier (three-unit and six-unit GB10 layouts, the TensorFold EXL3 test, single-card H200 layouts for models that fit better elsewhere) are on the leaderboards page.
If You Only Remember Four Lines
Buying one small appliance? Run Gemma-4-26B-A4B for anything a customer or auditor will read, and Nemotron-3.5-Lightning for high-volume drafting. Both tier S on one GB10 for different reasons.
Building a small pod? Four GB10 units on the RoCE fabric run the best combination in the fleet: GLM-5.3-Flash when the answer must be right, Qwen3.8-Flash-Next when the queue must keep moving. Quality tier S, throughput tier S.
Buying the workstation box? The DGX Station GB300 is where GLM-5.3-Flash shows its ceiling: a needle verified at a depth of 968,377 tokens and 38 of 38 tool calls in one quiet box, with gpt-oss-120b alongside for raw single-user speed at 390.7 tok/s.
Running a real production pool? Four H200s with GLM-5.3-Flash FP8 remain the overall leader: 5,245 tok/s at peak measured concurrency, 90 percent strict retrieval accuracy, zero measured hallucination rate. Nothing else we measured serves more users as accurately.
Whichever tier you land on, the model pages carry every number with its source file, and the methodology page explains exactly how we gate a result before we record it. If you want the same battery run against hardware you already own, that is a service we sell.
Pick the tier, then prove it on your own data. Test a model on your hardware Call Penny at 919-348-4912
Frequently Asked Questions
Straight answers from our own measurements. Every number traces to a stored result file listed on the model pages.
Are these tiers objective?
No, and we say so plainly. The tiers are Petronella Technology Group, Inc.'s judgment about which measured configuration serves private business use best at each hardware budget. The measurements underneath them are objective and reproducible; the weighting between speed, concurrency, verified context and retrieval accuracy is ours.
What does the S, A, B, C scale mean here?
S means the configuration we would deploy first for that budget, A means a strong choice with a trade-off you must understand, B means measured and beaten by a sibling for the same job, and C means we measured it and would not build it that way again. A model can be S on one box and C on another.
Why does Gemma-4-26B-A4B outrank bigger models on a single GB10?
Because at that budget the question is what one box can do well. Gemma-4-26B-A4B served 37 of 38 tool tasks, a verified context of 243,271 tokens and 62 percent strict grounded compliance, at 31.5 tok/s for one user and 866.4 tok/s across 64 users. Bigger checkpoints either did not fit or dropped to a few tokens per second.
Why is GLM-5.3-Flash not tier S on the two-unit GB10 pair?
Measured, not assumed: on two units in tensor parallel it managed 24.1 tok/s prose, 68.3 tok/s at its 6-user ceiling and 0.23 tok/J, the weakest pair layout we measured. The same model is tier S on the four-unit pod and the DGX Station GB300, where it gets the memory bandwidth it needs.
What is CMMC RAG accuracy and why does it matter for tiers?
It is our retrieval-grounded compliance battery: the model answers questions over a fixed document set and we score whether the answer cites the right control. For private business use it matters because a model that invents a citation is worse than no model. GLM-5.3-Flash and Qwen3.8-27B FP8 both scored 90 percent strict on four H200s; Nemotron-3.5-Lightning and gpt-oss-120b sit at the bottom of their tiers on this row.
Which model should I pick for 30 to 60 concurrent users on a budget?
The four-unit GB10 pod running Qwen3.8-Flash-Next: 648.7 tok/s across 64 concurrent users with the pod's best prefill. If the answers feed a compliance workflow, run GLM-5.3-Flash on the same four units instead and accept 258.6 tok/s at 32 users.
Is the DGX Station GB300 worth it over a GB10 pod?
For verified long context, yes: it found a needle at a verified depth of 968,377 tokens in one prompt, the deepest verified context in our fleet, and one user gets 145.9 tok/s of GLM-5.3-Flash. For serving many users per dollar, the pod wins. The box also has no published MSRP, so we budget against the reported launch figure.
Why do gpt-oss-120b and Qwen3.8-Flash-Next lose accuracy tiers on H200?
They do not lose speed; they lose the grounded-accuracy row. gpt-oss-120b scored 61 to 76 percent strict on the CMMC RAG battery against 90 for GLM-5.3-Flash and Qwen3.8-27B FP8 on the same pool, so we tier it A for volume work and reserve S for the models that answer compliance questions correctly.
What is the best single-user model you measured overall?
gpt-oss-120b on the DGX Station GB300 at 390.7 tok/s prose, with gpt-oss-20b at 549.2 tok/s if the task fits a smaller model. Qwen3.6-35B-A3B in BF16 is close behind at 335 tok/s on the same box.
Can a Strix Halo mini PC really run a 120B model?
Yes: gpt-oss-120b in MXFP4 ran at 41 tok/s on our ai11 box and about 40 tok/s on the g1a laptop, using unified memory. It is a single-user machine and prefill collapses on the wrong runtime, but the model genuinely runs.
How current are these tiers?
The measurements behind them run from September 23 to October 23, 2026, and each model page states its measured date and result files. We re-tier when a new measurement lands, and the methodology page documents the two results we have retracted after auditing our own work.
Can you run this battery on hardware we already own?
Yes, that is a standard engagement: we run the same gated battery against your hardware and your own retrieval set, then hand you the tier list for your fleet. Call Penny at 919-348-4912 or use the benchmark contact form to start.
Want Your Own Tier List?
These tiers came from our fleet. Yours will differ with your hardware, your documents and your users. We run the same measured battery on your equipment and hand you the placements with every number traceable to a stored result.