How well do local AI models work?
What our test runs have measured so far, to sort through: 11 results, for 9 models, on 2 computers. That isn't many yet, and we add to it as more runs happen. Each has a Score, and we don't pick a winner.
Scores from different computers or method versions aren't the same test, so compare them with care.
| Share | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| LFM3 1.7B | 95 (v1) | 57 of 64 | 22 of 22 | 20 of 21 | 38 of 39 | – | – | – | Apple M5 Pro, 24 GB | Nov 2, 2026 | 1 | Post |
| Gemma 2 9B | 90 (v1) | 60 of 64 | 20 of 22 | 20 of 21 | 32 of 39 | – | – | – | AMD RX 7900 XTX, Vulkan | Nov 2, 2026 | 1 | Post |
| Qwen 2.5 14B | 98 (v1) | 64 of 64 | 22 of 22 | 20 of 21 | 37 of 39 | – | – | – | AMD RX 7900 XTX, Vulkan | Nov 2, 2026 | 1 | Post |
| Qwen 2.5 32B | 99 (v1) | 63 of 64 | 22 of 22 | 21 of 21 | 38 of 39 | – | – | – | AMD RX 7900 XTX, Vulkan | Nov 2, 2026 | 1 | Post |
| Qwen 2.5 3B | 85 (v0) | 62 of 64 | 19 of 22 | 18 of 21 | 28 of 39 | – | – | – | AMD RX 7900 XTX, Vulkan | Oct 31, 2026 | 0 | Post |
| Qwen 3.5 4B | 96 (v0) | 61 of 63 | 22 of 22 | 18 of 21 | 39 of 39 | – | – | – | AMD RX 7900 XTX, Vulkan | Oct 31, 2026 | 0 | Post |
| Qwen 2.5 7B | 96 (v0) | 64 of 64 | 20 of 22 | 21 of 21 | 36 of 39 | – | – | – | AMD RX 7900 XTX, Vulkan | Oct 31, 2026 | 0 | Post |
| Qwen 3.5 9B | 96 (v0) | 60 of 63 | 21 of 22 | 20 of 21 | 38 of 39 | – | – | – | AMD RX 7900 XTX, Vulkan | Oct 31, 2026 | 0 | Post |
| Qwen 2.5 14B | 98 (v0) | 64 of 64 | 22 of 22 | 20 of 21 | 37 of 39 | – | – | – | AMD RX 7900 XTX, Vulkan | Oct 31, 2026 | 0 | Post |
| Qwen 3.8 27B | 98 (v0) | 62 of 63 | 22 of 22 | 20 of 21 | 38 of 39 | – | – | – | AMD RX 7900 XTX, Vulkan | Oct 31, 2026 | 0 | Post |
| Qwen 2.5 32B | 99 (v0) | 64 of 64 | 22 of 22 | 21 of 21 | 38 of 39 | – | – | – | AMD RX 7900 XTX, Vulkan | Oct 31, 2026 | 0 | Post |
Add yours
Have a computer we don't?
Every result helps people with hardware like yours.
Today this takes a checkout of Toskar, and it runs in a throwaway copy that leaves your everyday chats alone. How to add your results. We're working on a single line to paste on Mac, Linux, and Windows.
Questions
- Which local AI model is best?
- We don't say. These results show what we measured, on which computer, and on what date, so you can sort them by what matters to you. They come from few runs so far, and a model that does well on one set can do poorly on another.
- What does the Score mean?
- It's the plain average of the pass rates of the Quality, Topics, Files, and Reasoning sets, from 0 to 100, shown with the version of the method, such as 95 (v1). It's a score for these sets and not a verdict: they're easy enough that good models bunch near the top. Speed sits beside it and never counts toward it.
- What do reads and writes mean?
- Writes is how fast an answer appears once it starts. Reads is how fast a model takes in what you send, which is often the real wait in a long chat. First token is how long you wait after sending a long prompt. Newer results have these. Older ones weren't measured this way, so they show a dash.
- Will this model run on my computer?
- That comes down to memory. Our guide to what your computer can run has an estimate for each model, and these results show how models did on the computers we've tested.
- How is this measured?
- Each set is a fixed list of requests that a model answers through Toskar, with checks on what it did. The cases, checks, and thresholds are all public, and so is what we know is weak about them.
- Can I compare any two rows?
- With care. Scores from different computers aren't the same test, and neither are scores from different versions of the method. The filter shows one computer at a time.
Next
What will run on my computer?
Check your Mac or PC to see which models fit its memory.