Mistral just caught up: Mistral Large 4
I never thought I’d be saying this, but out of absolutely nowhere, Mistral dropped a new model that really changed its position. The model is called Mistral Large 4 (nicknamed Le Chonk) and to show how significant this is, their last model scored 9 and this one scored 38, which is a massive jump of 29 points on the Artificial Analysis Intelligence Index. It’s level in intelligence with GPT-6 Luna (max), so it’s nowhere near the current frontier of 58 (Opus 5.5), but it doesn’t need to be. Mistral is the highest-scoring open model outside China.
- Claude Opus 5.5 (max)US58
- Claude Sonnet 5.5 (max)US56
- GPT-6 Astra (max)US53
- Gemini 4 Argent (high)US53
- GPT-6.1 Sol (max)US52
- Muse Spark 1.3 (max)US48
- Grok 4.7 (xhigh)US46
- Qwen3.8 MaxChina45
- GLM-5.3 (max)China45
- Gemini 3.8 Flash (high)US41
- DeepSeek V4.1 Flash (max)China39
- GPT-6 Luna (max)US38
- Mistral Large 4 PreviewEurope38
Artificial Analysis Intelligence Index v4.3.2, checked 7 October 2026.
Just like other models around this size, it has a 1 million token context window, and its weights are planned to be released by the end of October, so anyone can download it and run it themselves, assuming you have a mini datacentre in your living room.
The only reference I even had of Mistral was around 2025, when I used it briefly and it fucking sucked, but it’s good that they’ve gone from kind of a laughing stock to a serious contender. They were praised for their smaller models and their open nature, but for me at the time, I formed a negative opinion of them because their models were just so weak compared to the frontier back then. I’d look at it differently now, as I’ve come to appreciate open models a bit more.
I like that Europe has its own model independent of America or China. I enjoy seeing the competition between these now three powers, although Europe will probably lag behind (although maybe not, since they’ve already made this big leap).
I wish the UK had its own model, but we don’t, and I think it’s far too late by this point to ever have one. One company, Cosine, is trying, with a model aimed for the end of 2026. We’ll see how that goes.