Mistral Opens Preview of Large 4, a Trillion-Parameter Open-Weight Model

Mistral AI has opened a public preview of Mistral Large 4, a model with about one trillion total parameters that the company's fans have nicknamed “Le Chonk.” Announced on October 6, the model is available through Mistral's API, and the company says downloadable weights will follow by the end of October. Large 4 is considerably bigger than its predecessor, Large 3, which had 675 billion parameters according to Decrypt. Mistral's benchmark comparisons focus on Chinese open-weight models.

The architecture and training.

The model uses a mixture-of-experts design, so only part of it runs for each token. The coverage gives two figures for the active portion. Decrypt and The Register report about 49 billion parameters, while Developers Digest uses 52 billion, which it takes from Mistral's current documentation. The model is natively multimodal and accepts text and images as input. Developers Digest and The Register both report that Mistral trained it on roughly 3,800 Nvidia Grace Blackwell GPUs in its European data centres.

Pricing and the context window.

Mistral's model card lists a context window of one million tokens, according to Let's Data Science. Developers Digest notes that the served preview runs with a window nearer 524,000 tokens, and it advises planning around that smaller number. List pricing is $1.36 per million input tokens and $4.18 per million output tokens, according to Decrypt. Developers Digest reports a temporary sale price of $0.68 and $2.09, with no end date announced, and suggests budgeting at list price.

What the benchmarks show.

Most of the benchmark figures so far come from Mistral or from leaderboards it selected. On a blind human rating from Surge AI, Large 4 scored 3.74 out of 5, against 4.22 for Anthropic's Claude Opus 5, Decrypt reports. On the Artificial Analysis Intelligence Index, Developers Digest reports a score of 38, compared with a median of 26 for models in the same price class. The Register says the index places the preview between DeepSeek V4.1 Flash and OpenAI's GPT6 Luna, and it describes that position as a clear disadvantage against the flagship models from OpenAI and Anthropic.

Safety testing and the weights.

Mistral says it is continuing red-team testing with selected cybersecurity partners and public-sector organisations before the release is complete, according to Let's Data Science. Decrypt quotes Guillaume Lample, the company's chief scientist, calling it “by far the strongest open-weight model from the US or Europe.” Developers Digest reports that Lample said the reinforcement learning run is still under way, so the final downloadable version could differ from the preview. None of the coverage states the licence terms.

What is still unknown.

The reports disagree on active parameters and on the usable context length, and the licence has not been published in the coverage. Pricing may change once the sale ends. The benchmark numbers mostly come from the company and from third-party leaderboards that it chose, so independent tests of the final weights will matter more than the preview scores. Teams in Europe and Asia who plan to download the weights on release day can check the local time with the free world clock at /en/world-clock.