Mistral Large 4: the sovereign open model with 1 trillion parameters and a license to read first
Direct answer
On October 29, 2026, Mistral AI published the weights of Mistral Large 4, nicknamed Le Chonk, three weeks after its public preview on October 6. It is a 1-trillion-parameter model in total, with about 49 billion active per token, a 1-million-token context window and text and image input. The license is Mistral's Ecosystem Open-Weight License, a modified MIT license: free use, modification, fine-tuning and commercialization, with one condition that changes the business decision.
The license, with no surprises later
- Running it in-house carries no license fee: a bank, ministry or manufacturer running the model for its own staff pays no license, per license coverage.
- Offering it as a service needs a separate license: a business operating a model-as-a-service offering needs a separate commercial license from Mistral before using it commercially.
- The grey line is wrapped SaaS: coverage notes the announcement does not define precisely where a SaaS product using the model ends and a model service begins. Before building a business on that line, the license is read with counsel, not from the headline.
- Downloaded weights do not update themselves: the published version reflects three additional weeks of reinforcement learning after preview, and weights downloaded today will not receive future changes without downloading a new version.
Hardware decides the decision
- Four GPUs as a practical floor: the compressed version runs on a server with four Nvidia B200 or B300 GPUs, per coverage, and Mistral's deployment examples use eight GPUs. A trillion parameters is about 2.2 terabytes of weights at BF16 precision and about 1.1 terabytes at 8-bit.
- The API route already has a price: through Mistral's API, the model costs $0.70 per million input tokens and $2.10 per million output tokens, with Microsoft Foundry as the first cloud carrying it and Nvidia offering a NIM container.
- The focus is code and cyber defense: Mistral reports 48% on Artificial Analysis's Coding Agent Index, using 47% fewer tokens per task than GLM-5.3, and 51% on the same firm's Cyber Index, with 96.6% refusal of malicious cyber requests in its tests. Those are company figures, with no independent verification published.
- The aggressive version stays restricted: a more aggressive penetration-testing variant stays reserved for vetted security firms, such as Orange Cyberdefense, with early access.
- Distilled models are coming: Mistral said it will release distilled models in the coming months, including one that runs on a laptop. The sovereign route for small teams may be waiting for that distilled model, not buying the eight-GPU server.
Note: this article summarizes Mistral AI's October 29, 2026 announcement and TechCrunch, The Next Web and Skimwire coverage, plus the Hugging Face model card (Mistral-Large-4-1T-A52B). Performance and hardware figures are company-reported, with no independent verification published as of that date.
Want AI that works for your business?
Message us on WhatsApp and let's talk about automation and AI agents for your business.
WhatsApp inicIARelated articles
- How to choose an AI vendor: 7 questions before signing
- What an AI agent is and how a real business uses one
- Sema4 3.0 promises 80% fewer tokens with forensic audit
Frequently asked questions
What did Mistral publish on October 29, 2026?
The weights of Mistral Large 4, a 1-trillion-parameter model with about 49 billion active per token, a 1-million-token context window and text and image input, available on Hugging Face.
Is Mistral Large 4 free to use?
Running it for in-house use carries no license fee under its modified MIT open-weight license. Offering it as a model service requires a separate commercial license from Mistral.
What hardware does it need?
The compressed version needs a server with four Nvidia B200 or B300 GPUs, and Mistral's deployment examples use eight GPUs. It is not a laptop model, except for the distilled models announced for the coming months.
What does it cost via API?
$0.70 per million input tokens and $2.10 per million output tokens through Mistral's API, per launch coverage.
