Mistral Large 4 release

Import AI placed Large 4 at 62 on DeepSWE 1.1 on October 12, behind Kimi K3's 68 and just above GLM-5.3's 61, and characterized it as the expected strongest Western open-weight model at release but weaker than Chinese models on most benchmarks.

Mistral separately published a Dense 200 visual grounding score of 42%, claiming it surpasses frontier closed models, and an NVIDIA blog post described Large 4 as under private preview and a foundation for customization across cyber defense, manufacturing, and finance.

Mistral Large 4 referenced in AI roundup as a one-trillion-parameter open-weight model nicknamed 'le Chonk.'

Mistral released Large 4 on October 6 as a public preview, a one-trillion-parameter mixture-of-experts model described as natively multimodal and built entirely in Europe.

Claims in coverage that it beats Claude, ChatGPT, and Gemini are not supported by Mistral's own benchmarks, which rank Large 4 behind Claude Opus 5 in blind coding evaluation. Open weights are expected by end of October after a safety review, with license terms unconfirmed.

Simon Willison described Mistral Large 4, the one-trillion-parameter model Mistral previewed via API on October 6, as putting Mistral 'back in the game' but 'maybe about 6 months behind the frontier' and 'certainly not a Fable-class model'.

On Artificial Analysis benchmarks it scores 38, just behind DeepSeek 4.1 Flash and well above Mistral Large 3's score of 9. Cofounder Guillaume Lample positioned the model around specialized domains, noting 'There are a lot of areas where the other labs will not focus that much'.