Model data

What actually fits in an office.

Benchmark tables are everywhere and almost none of them answer the only question that matters here: whether a model runs on hardware you could put under a desk, and what you give up by choosing it.

Every model, both kinds

Open-weight models you could run in your own building, and the hosted ones you can only rent. One table, because the real decision is between them rather than within either. Rebuilt every morning from the sources listed at the bottom.

Showing 10 of 40 · 102 GB usable on the selected machine

RunsSpeedLicence
Llama-3.1-8B-Instructmeta-llamaYour machine5.9 GBmeasured84tok/sllama3.1
gpt-oss-120bmy pickopenaiYour machineone command75.4 GBmeasured116tok/sapache-2.0
gpt-oss-20bmy pickopenaiYour machineone command13.9 GBmeasured161tok/sapache-2.0
Meta-Llama-3-8B-Instructmeta-llamaYour machine4.8 GBestimated84tok/sllama3
Llama-3.2-3B-Instructmeta-llamaYour machine2.4 GBmeasured211tok/sllama3.2
Llama-3.1-8Bmeta-llamaYour machine4.8 GBestimated84tok/sllama3.1
GLM-4.7-Flashzai-orgYour machineone command22 GBmeasured147tok/smit
Qwen2.5-7B-InstructQwenYour machine4.6 GBestimated89tok/sapache-2.0
gemma-2-2b-itgoogleYour machine1.6 GBestimated260tok/sgemma
Qwen3-8BQwenYour machineone command6 GBmeasured82tok/sapache-2.0

Speed is estimated, not measured. Memory bandwidth divided by the weights actually read per token, times an efficiency factor of 0.55, the formula published by llmfit (MIT). For a mixture-of-experts model it uses the active parameters rather than the total, which is why a 117B model can outrun an 8B one. Treat it as an upper bound, and ask for a measured number before you rely on it.

A blank rating means Arena has not rated that model, not that it scored badly. Ratings are joined by exact name only; a looser match produced confident nonsense, so it was taken out.

Rebuilt 2026-08-31 fromHugging Face Hub APIOpenRouterOllama libraryUnsloth GGUFsllmfit (speed formula, MIT)LobeHub iconsArena, CC-BY 4.0

On a 128 GB machine

Snapshot · 2026-07-28

ModelSizeFootprintSpeedComparable to
gpt-oss-120bThe workhorse, and the reason a 128 GB office machine is viable at all. Only a fraction of its parameters activate per token, so it runs fast on hardware where a dense model of the same size would crawl. Quantisation-aware trained, so the 4-bit checkpoint is the reference rather than a lossy copy.120B (MoE, ~5B active)~63 GB · MXFP4Fits34–55 tok/s≈ o4-mini
Qwen3-30B-A3BWhere responsiveness matters more than depth. Strong at pulling structured data out of documents and at answering from retrieved context, which is most of what an office actually asks.30B (MoE, ~3B active)~18 GB · Q4Fits~100 tok/s≈ GPT-4o class on extraction
Qwen3-235B-A22BOnly fits by quantising hard enough to erode the reasoning advantage you wanted it for. A good illustration of why parameter count is the wrong thing to shop on.235B (MoE, ~22B active)~142 GB at Q4Barelydegraded≈ DeepSeek-R1 tier
Dense 70B modelsThe cautionary entry. It loads, and then produces text more slowly than you can read. These machines are limited by memory bandwidth, not capacity, which is exactly the trap a spec sheet leads you into.70B (dense)~40 GB at Q4Barely~2–3 tok/s
GLM / DeepSeek frontierGenuinely competitive with closed frontier models, and completely out of reach of an office machine. Available as a hosted endpoint if you want their quality with your own adapter, at the cost of the data leaving the building.355B – 600B+datacentreNohosted only≈ near-frontier

The single most useful thing on this page: never buy a machine on parameter count. These boxes are limited by memory bandwidth, not capacity: a dense 70B will load happily and then run at reading speed. Ask for measured tokens per second on the specific model you’ll actually be given.

The gap, measured

Epoch AI · runs to 2026-08-02

The best open-weights model that fits a 128 GB office machine, against the best score any model has posted, on the same benchmark. Not an average and not an index. Per task, because the answer changes completely depending on what you are asking for.

  • GPQA Diamond

    18.8 points behind

    Graduate-level science questions.

    Runs in your office · openai/gpt-oss-120b_high75.8%
    Best anywhere · GPT-5.4 Pro (xhigh)94.6%
  • MATH Level 5

    8.2 points behind

    The hardest tier of competition mathematics.

    Runs in your office · DeepSeek-R1-Distill-Llama-70B89.9%
    Best anywhere · GPT-5 (high)98.1%
  • SimpleQA Verified

    63.4 points behind

    Short factual questions. Measures making things up.

    Runs in your office · openai/gpt-oss-120b_high13.9%
    Best anywhere · Gemini 3.1 Pro Preview77.3%
  • AIME (mock)

    11.1 points behind

    Olympiad-level mathematics.

    Runs in your office · openai/gpt-oss-120b_high88.9%
    Best anywhere · GPT-5.6 Sol (max)100.0%

What the widest gap tells you

SimpleQA Verified is the worst result on this page: 63 points, against single figures on hard mathematics. It is a test of recalling facts without help, and a model small enough to sit under a desk is bad at it in a way no amount of tuning fixes.

That is an argument for the architecture rather than against the machine. A local model should not be answering from memory. It should be reading your documents and answering from those, which is retrieval, and retrieval is where the gap above largely stops mattering.

Benchmark data from Epoch AI, used under CC-BY 4.0, drawn from 1,107 evaluation runs. “Runs in your office” means an open-weights model of 200B parameters or fewer, which fits 128 GB at 4-bit. That is a test of whether it loads, not of whether it is fast enough to use. Those are different questions, and the second one decides whether anyone keeps opening it.

What people actually prefer

Arena · CC-BY

Benchmarks measure whether a model can do a task. These measure whether people preferred its answer, blind, against another model given the same prompt. Millions of votes, and a confidence interval on every row so you can see when a ranking is real and when it is noise.

Blind head-to-head preference on everyday prompts. The closest thing that exists to a measure of whether people actually like using a model.

395 ranked · 3.6M votes · published 2026-08-27

RankModelLicenceRating
1claude-opus-5-maxanthropicProprietary1505
2claude-opus-5-highanthropicProprietary1504
3claude-opus-4-6-highanthropicProprietary1503
4claude-opus-4-6anthropicProprietary1497
5claude-fable-5anthropicProprietary1495
6gemini-3.7-flash-highgoogleProprietary1490
7claude-opus-4-7-highanthropicProprietary1490
8muse-spark-1.2 (xHigh)metaProprietary1488
9gemini-3.5-flash-highgoogleProprietary1483
10claude-opus-4-7anthropicProprietary1483

Open weights is not the same as runs on your box

The licence column says whether you are allowed to download and run a model. It says nothing about whether you can afford to. Several of the open models ranked here are large enough to need a rack of server GPUs, so read this as a measure of quality and of freedom, not of whether it fits under a desk. The question of what fits 128 GB is answered further up this page, where the data carries parameter counts and this data does not.

Leaderboard data from Arena, used under CC-BY 4.0 via their published dataset. Ratings are Bradley-Terry scores from blind pairwise votes; the agent board instead reports an estimated causal effect from a randomised trial over real sessions, described in their methodology note. The two scales are not comparable and are never shown together.

The full index

A second, much larger index covers several hundred models. Its licence permits internal use but not republishing, so it sits in the client area rather than here.

If you are weighing a purchase and want to see where a specific model lands, ask me and I will walk you through it directly. That is a better conversation than a table anyway.

The same list, as a picture

One line, one question. Everything to the left of it runs on a machine you can put under a desk; everything to the right needs a rack, a cloud account, or both. Most of the models people name in meetings are on the wrong side of it.

Memory needed to serve · Mac Studio M5 Max

102 GB usable of 128

  • Motif-3189
  • Hy3179
  • DeepSeek-V4-Flash166 · 138 file
  • DeepSeek-V4-Flash-0731164 · 137 file
  • Laguna-S-2.188 · 73 file
  • Ling-3.0-flash77
  • gpt-oss-120b75 · 63 file
  • Karnak-40B-v1.024
  • GLM-4.7-Flash22 · 18 file
  • Ornith-1.5-35B-A3B22
  • Apodex-1.1-mini22
  • NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF1619
  • phonellm-alpha-119
  • granite-4.2-30b18
  • Qwen3.8-27B-OBLITERATED17
  • Huihui-Qwen3.8-27B-abliterated-NVFP416
  • gpt-oss-20b14 · 12 file
  • Qwen3.6-35B-A3B-NVFP411
  • NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP411
  • LFM2.5-8B-A1B6 · 5 file
  • Qwen3-8B6 · 5 file
  • Llama-3.1-8B-Instruct6 · 5 file
  • Ornith-1.5-9B6
  • Qwen3.8-9B-Distill6
  • Ornith-1.5-9B-OBLITERATED6
  • Qwen2.5-Coder-7B-Instruct6 · 5 file
  • granite-4.2-8b5
  • zeta-2.15
  • Meta-Llama-3-8B-Instruct5
  • Llama-3.1-8B5
  • Ling-3.0-tiny5
  • Qwen2.5-7B-Instruct5
  • Qwen3.8-27B-Escha-W24
  • Qwen3.8-27B-MTPLX-Optimized-Speed4
  • Nanbeige4.2-3B3
  • Llama-3.2-3B-Instruct2 · 2 file
  • granite-4.2-3b2
  • gemma-2-2b-it2
  • LFM2.5-2.6B2
  • Qwen3.8-27B-DFlash21
  • Qwen3.8-27B-DFlash21
  • DFM-Mimir1
  • GLM-5.3-Flash-DFlash21
  • MiniCPM5-1B1

Bars are gigabytes of memory needed to serve the model: 4-bit weights plus about 20% for KV cache and runtime. The second figure is the download, which is smaller and is not the number that decides whether it runs. Dark bars clear the line on a Mac Studio M5 Max; grey ones do not. 7 larger models are left off the chart entirely, all of them past 204 GB.

Where local is enough, and where it isn’t

Split by task rather than by an overall score, because the average hides the thing you need to know.

  • Summarising a document

    Local is fine

    No practical difference on the kind of documents a firm produces.

  • Pulling structured data out of a form

    Local is fine

    Reliable. This is the single best-value use of a local model.

  • Answering from your own files

    Local is fine

    Where retrieval does the work, the model only has to read well.

  • Drafting in a house style

    Local is fine

    With a trained adapter, often better than a frontier model without one.

  • Bounded tool use

    Local is fine

    Three or four dependent steps is comfortable.

  • Reasoning across 100+ pages at once

    Frontier ahead

    The clearest weakness. Quality falls off well before the advertised context limit.

  • Long autonomous agent runs

    Frontier ahead

    Per-step errors compound. Thirty dependent steps is not reliable locally.

  • Complex multi-file coding

    Frontier ahead

    A wide margin, and the honest reason to keep paying for a frontier API.

About these numbers

Three different things are on this page. The table at the top is a dated, hand-maintained snapshot. The benchmark comparison is rebuilt from Epoch AI’s published runs and carries the date of the most recent one. The leaderboards are rebuilt from Arena’s published votes and carry their publication date. Anything presented as current without a date is worth distrusting, here or anywhere else.

Both external sources are used under CC-BY, which is the reason they are the ones on this page. There are better-known indexes whose licences permit internal use but not republishing, and one of those sits behind the client sign-in below rather than being quietly reprinted here.

Both are baked into the page when it is built, never fetched from your browser. A privacy argument that made your machine call a third party to render itself would not be much of an argument.

Speeds are drawn from published benchmarks on comparable hardware, not measured by me on a box in front of a client. When I’ve measured them myself, that’s what will be here instead.

The longer argument →