Who makes it
The AI Bench is written and maintained independently. It has no ads, no affiliate links and no sign-up. The checker works out everything in your browser, and we do not store your hardware details. We count page visits with Cloudflare’s cookie-free analytics.
The site was rebuilt from scratch in October 2026 around one tool, with fewer and deeper pages.
How AI is used here
We use AI assistants to collect sources, write code and draft text. AI is not trusted for facts. Every text model size comes from the actual download files on Hugging Face; image and speech memory comes from the model card or the file sizes; most chip specs come from the maker’s page, and the rest from TechPowerUp’s database; every community speed has a linked source, and our own runs list their exact settings. Automated tests check the data on every change — for example, that our speed estimate stays centred on about 770 measured runs.
How we work out memory
A model needs memory for three things:
- The model file — the exact size of the download (for example the Q4_K_M file).
- The conversation (KV cache) — worked out from each model’s design: layers × heads × head size × 2 (keys and values) × 2 bytes per number, for the number of tokens you choose. If a model does not publish this, we assume 15% of the file size per 32K tokens.
- The app — a fixed 1 GB.
What you can use: all of a graphics card’s memory; on a Mac, the limit macOS sets by default — about two-thirds of memory on 24–32 GB Macs, about 56 GB on a 64 GB Mac and about 120 GB on a 128 GB Mac (values reported by llama.cpp on real Macs) — and more after raising the limit with one command; 70% of RAM when there is no graphics card. If a model is too big for the graphics card, we check whether it fits in the card plus 70% of your RAM — it runs, but slowly.
Labels: Runs great = at least 35% spare, Runs well = at least 10% spare, Just fits = fits with little room, Fits with a tweak = only after raising the Mac limit.
How we estimate speed
To write each word (token), the computer reads the active part of the model, plus the conversation, from memory once. So the time per token is about:
bytes read ÷ (efficiency × memory bandwidth) + a small fixed cost
We fitted the numbers on about 774 public measurements: graphics cards and other PCs with llama.cpp, and Macs with MLX (what Ollama and LM Studio use on a Mac). On graphics cards: efficiency 0.8 with a 0.8 ms fixed cost for normal (“dense”) models, and 0.59 with 2.6 ms for mixture-of-experts models. On Macs: 0.89 with 1 ms, and 0.9 with 6 ms — mixture-of-experts models on a Mac are held back more by the fixed cost than by memory speed. Apple’s Ultra chips reach less of their bandwidth: we count 78% of it for dense models and 50% for mixture-of-experts. Two cards measure well below their bandwidth, so we scale them down: RX 7900 XTX and RTX A6000. llama.cpp on a Mac is about a quarter slower than MLX; the low end of the range covers it. Today 49 of our 52 older calibration runs fall inside the range, and the middle of the range is typically 9% away from the measured speed. See them on the measurements page.
Limits: the estimate is for writing the answer, not for reading a long prompt first (Macs are much slower at that). App versions matter too: in one public test, the same Mac got about a third faster from a llama.cpp update alone.
How models are ranked
For each task we rank models that fit your machine by a quality score, then push down models that only just fit or need a tweak, and models that are too slow for the task. Chat and reading images feel fine at about 8 tokens a second; long documents need about 15; coding and agents write much more, so they need about 20. For the same reason we may pick the faster 4-bit file over the 8-bit one for coding. When Artificial Analysis scores a model at several thinking efforts, we use the normal (Medium) one, like the other models.
- Text models: the Artificial Analysis Intelligence Index (v4.3.2, read on 10 October 2026). 28 sizes are measured by Artificial Analysis; 38 are their estimates, partly based on the makers’ own claims (marked on each model page); 7 sizes they do not list use the score of their base model or a close peer. For coding, long documents, agents and reading images, half the score comes from tests of that task (coding: Terminal-Bench Hard, SciCode and SWE-rebench real GitHub fixes, compared only within the same SWE-rebench task window; long documents: AA-LCR; agents: τ²-Bench and Terminal-Bench Hard; images: MMMU-Pro), scaled to match the index. For chat and writing, three quarters comes from people’s votes on LMArena (overall and creative writing). A model without a test gets its maker’s usual gap on that test, or keeps its overall score if we have too few of the maker’s models. An older model whose direct successor also fits your machine is listed below every current model.
- Image and speech models: no shared quality index covers them, so we rank by downloads in the last 30 days, how new they are, and whether business use is allowed.
We show one size per model family: the best one for your machine.
What we track
- 52 model families and 107 model sizes, with file sizes checked on 10 October 2026 and Ollama names re-checked on 13 October 2026.
- 105 graphics cards and chips, with specs from maker pages.
- 63 measured speeds with sources.
- Hardware prices with the date we checked them. Prices moved a lot in 2026.
Mistakes and corrections
When we get something wrong, we fix it and say so on the page. The latest: two benchmark rows on our old site came from a different NVIDIA device (details on the measurements page).
Found a mistake? Message @TheAIBench on X or open an issue on GitHub.