Mistral Large 4 Preview: 1T-Param Model Available, Weights by End of October

Mistral released a preview of Mistral Large 4 on October 6, 2026, a natively multimodal mixture-of-experts model with 1 trillion total parameters and 49 billion active parameters, trained on 3,800 NVIDIA Grace Blackwell GPUs. The preview is accessible via Mistral Studio and the Mistral API using model ID `mistral-large-4`. Open weights are expected by end of October 2026, pending a safety review, though a reported release date of October 27 came from third-party coverage rather than Mistral's official announcement. Mistral describes the model as its largest and most capable to date. The API currently supports two reasoning levels, 'none' and 'high'. Cofounder Guillaume Lample said there are many specialized domains where other labs will not focus, and the model targets coding, cyberdefense, manufacturing, finance, and electrical engineering. On Artificial Analysis benchmarks, the model scores 38, placing it just behind DeepSeek 4.1 Flash and well above Mistral Large 3, which scored 9. Simon Willison wrote that Mistral is 'back in the game' but described the model as 'maybe about 6 months behind the frontier' and not at the level of frontier models. Mistral's own blind coding evaluation ranks it behind Claude Opus 5, and claims circulating in coverage that it beats Claude, ChatGPT, and Gemini are not supported by Mistral's benchmark data. The license terms for the open-weight release have not been confirmed.