Mistral Releases Large 4 as an Open-Weight Model, With About 1 Trillion Parameters

Mistral has released the weights for Mistral Large 4, nicknamed Le Chonk, after a hosted public preview earlier in October. Mistral describes it as a natively multimodal model that combines instruct and reasoning modes in a single mixture-of-experts design, with around one trillion parameters in total. The weights are on Hugging Face for self-hosting with vLLM or SGLang, and there is an NVIDIA NIM container.

The licence is the detail to read. It is the Mistral Ecosystem Open-Weight licence, described as a modified MIT licence. It allows free use, modification, fine-tuning and commercial use, but a business offering the model as a service to others needs a separate commercial licence from Mistral. For most studios and product teams that is fine. For a platform reselling access, it is not.

On price, Mistral lists $0.70 per million input tokens, $0.07 for cached input and $2.10 per million output tokens through its API. It is also available through Mistral Studio, OpenRouter, Microsoft Foundry and NVIDIA-hosted endpoints. There are three reasoning modes, low, medium and high, with high as the default.

The benchmark claims are Mistral's own. It reports 41 percent on Terminal-Bench 4.0, up 13 points from the preview, and says it is preferred in a blind human coding comparison by Surge Code over several open rivals, ranking behind only Claude Opus 5.5. It also reports a 96.6 percent average refusal rate on malicious cyber requests. Vendor-reported results are a guide to where to test, not proof it will suit your work.

Why does it matter to a creative studio? A trillion-parameter model is not something you run on a workstation, so most of us will use it through a hosted endpoint. The value of open weights is choice. You can pick a European-trained model, move providers without rewriting your prompts, and keep sensitive client material on infrastructure you control if you need to.

If you build with AI, add it to your test set. Run the same five real tasks through it that you run through your current model, and compare the results rather than the leaderboard.