Mistral Large 4 Review

Mistral Large 4 went into public preview on October 6, 2026, and for now it lives only in the preview API on Mistral Studio. Mistral describes a hybrid instruct and reasoning mixture-of-experts model with about 1 trillion parameters in total and 52 billion active per token, natively multimodal input, more than 160 languages and, in its documentation, a 1 million token context. It promises open weights by the end of October, along with the architecture details and post-training methods. The model card lists $1.36 in and $4.18 out per million tokens, and no license for the weights has been named.

The numbers Mistral publishes put it near the top of the open field rather than the top of the market. It scores 61.7 percent on DeepSWE v1.1, 28.3 percent on Terminal-Bench 4 and 59.9 percent on AutomationBench, ahead of Kimi K3 and DeepSeek V4 Pro on Mistral's chart, and it placed second of five in a blind human coding evaluation run with Surge AI, at 3.74 against Claude Opus 5's 4.22.

Frontier closed models score far higher on the same agentic tests: Claude Sonnet 5.5 posts 70.6 percent on Terminal-Bench 4.0. Its loudest result is in cybersecurity, 82 percent on a reproduce and patch test and a top five place on the Artificial Analysis Cyber Index, but Mistral itself notes that rivals such as Opus 5.5 score near zero there because they refuse the tasks.

Everything here is Mistral's own reporting, on a model whose reinforcement learning run is, in its words, still in flight. The price is low for its size, and once the weights arrive it will be the most capable model a European company can run on its own hardware. Until then it is a preview with one way in, which holds ease of use at a C, and the capability axes sit at B. It is a B, due a re-review when the weights land.