Most trillion-parameter models live behind an API key and a pricing page. Mistral just put one on Hugging Face. On October 29 the French AI lab released open weights for Mistral Large 4, which it has nicknamed "Le Chonk," and for once the nickname is fair.
One Trillion Parameters, 52 Billion on Shift
Large 4 is a mixture-of-experts model with 1 trillion total parameters, 52 billion of them active at a time, per Mistral's preview announcement earlier this month. It handles instruction-following, reasoning, and agentic work in a single model, accepts multimodal input, and offers three reasoning-effort settings (low, medium, and high) so developers can decide how hard it thinks on each request. Mistral says it trained the model from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its European data centers.
The weights ship under what Mistral calls the Mistral Ecosystem Open-Weight license, a modified MIT license. You can use, fine-tune, and commercialize it freely, with one catch: companies selling it as a hosted model service need a separate commercial license. If you'd rather not run it yourself, the API costs $0.70 per million input tokens and $2.10 per million output tokens, and it's also available through OpenRouter, NVIDIA's hosted endpoints, and Microsoft Foundry.
Open Weights, Real Choices
Mistral's own benchmarks put Large 4 at or near the top of the open-weight pack on coding-agent tests. Those numbers are self-reported, so wait for independent testing before crowning anything. For businesses, the bigger story is control. A model you can run on your own infrastructure means sensitive data never has to leave the building, which matters a great deal in healthcare, finance, legal, and anywhere a compliance officer has strong opinions.
A reality check, though: "you can download it" and "you can run it" mean very different things when the download is a trillion parameters. Self-hosting this one takes serious GPU hardware, not the spare server in the supply closet. Mistral also earns points for candor. Its release notes say a partner saw potential for the model to give real-world uplift to cyberattacks, and the company says it's tightening guardrails in response.
Large 4 is a genuine new option for companies that want frontier-class AI on their own terms. Whether it's the right option comes down to your data, your budget, and how many GPUs you're willing to buy.
If open-weight models have your team rethinking where your AI should live, our free AI implementation guide covers the hosting, privacy, and cost questions to settle before you commit.
Source: Mistral AI