Independent developer Simon Willison says Mistral Large 4 is "certainly not a Fable-class model," but that it puts Mistral "back to being maybe about 6 months behind the frontier"1. He called the preview, released on October 6, a huge improvement on Mistral Large 3 from last December.
Willison tested the preview through Mistral's API, which offers only two reasoning levels, "none" and "high". In his pelican-riding-a-bicycle drawing test, the "high" setting drew what he judged the better pelican while using fewer output tokens: 2,717, against 3,275 for "none".
He also pointed to Artificial Analysis, where Mistral Large 4 scores 38, just behind DeepSeek's 552-billion-parameter V4.1 Flash and up from 9 for Mistral Large 3. Artificial Analysis has called it the most intelligent model from outside the U.S. and China[1].
Earlier the same day, Willison took up a Hacker News commenter's joke that "frontier models are tested with an armadillo in fishnet tights jaywalking on Mars," asking Mistral Large 4, Claude Opus 5.5, GPT-6.1 Sol and Gemini 3.8 Flash for an SVG of that scene at each model's default reasoning level, and published the four results2. He has also released version 0.16 of his llm-mistral plugin, which adds support for reasoning models such as Mistral Large 43.
Mistral Large 4 is a 1.05-trillion-parameter mixture-of-experts model with 49 billion active parameters[1], trained on Mistral's own cluster of 3,800 Nvidia Grace Blackwell GPUs. The preview is available through Mistral's API, and the open weights are scheduled for October 27[1].