Mistral Large 4 weights arrive with a paid licence for model hosts

Mistral Large 4 weights arrive with a paid licence for model hosts

29 October 2026: Mistral AI has published the weights of Mistral Large 4, its one-trillion-parameter mixture-of-experts model, on Hugging Face, meeting the end-of-October date it set when the public preview opened on 6 October. The weights come under a new Mistral Ecosystem Open-Weight License: free for internal use, but any business that offers the model to others as a service must sign a separate commercial licence with Mistral before using it commercially.

A modified MIT licence with one large exception

Mistral describes the licence as a modified MIT licence, and most of it reads like one: use, modification, fine-tuning and commercial use are permitted, with the copyright notice kept in copies of the model. Section 2 is the change. If the licensee or any of its affiliates "operates a Model as a Service business", it must enter into a separate commercial licence with Mistral "before using the Model or its derivative works for any commercial purpose".

The licence defines model as a service as giving a third party access to the model or a derivative for inference or post-training, through an API or a hosted endpoint, in a way that lets that third party exercise "meaningful control over the inputs, parameters, or training data". Two cases are carved out of that definition: end-user products where the model sits inside specific features, and simply relaying requests to a model someone else hosts. Section 3 exempts internal use, meaning any use that does not make the model available to third parties, and use through Mistral's own products. The licence names no revenue threshold.

Hugging Face tags the repository with an "other" licence rather than a standard open-source one, which is accurate: a licence that requires some commercial users to negotiate is not an open-source licence in the OSI sense, whatever the weights' availability.

Who needs to read Section 2

The split falls along a line most enterprise buyers will recognise:

• A company running Large 4 on its own GPUs for its staff, or inside its own application, is covered by the internal-use exemption.
• A software vendor whose product calls the model behind a fixed feature, such as a summarise button, falls under the end-user product carve-out.
• A cloud provider, inference marketplace or reseller that exposes the model through an API where customers choose prompts and parameters is a model-as-a-service business and needs Mistral's commercial terms first.
• A systems integrator that hosts a fine-tuned copy for a client sits closest to the line, because the client may control the inputs. That case is worth putting to Mistral in writing before a contract is signed.

Mistral itself lists distribution through Microsoft Foundry, NVIDIA's hosted endpoints and NIM container, OpenRouter and its own API, which suggests the larger hosts already have agreements in place.

What the model is and what it costs on Mistral's API

The model card gives 52 billion active parameters per token out of about one trillion, a 1.6 billion parameter vision encoder and a context window of one million tokens. Mistral says it was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European data centres. Compared with the preview, Mistral says it added three weeks of reinforcement learning, replaced the "none" reasoning mode with low, medium and high modes, and changed its guardrails after seeing live cyber-related traffic during the preview.

Mistral's documentation lists API prices of $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens. Self-hosting is supported with vLLM and SGLang. The benchmark results in the announcement, including 48% on Artificial Analysis's Coding Agent Index, are Mistral's own reporting of those indices.

Jeff Boudier, VP Product at Hugging Face, is quoted in the announcement: "AI sovereignty isn't about choosing which API to call."

Where it sits in our catalogue

Our AI model category lists 143 models, among them Mistral Large 4, Mistral Large 2 and Mistral's Voxtral 3. Mistral has not said whether earlier models released under Apache 2.0 will move to the new licence; the licence on the Large 4 repository applies to Large 4 and its derivatives.

Company profile on TrustList: Mistral AI

Related on TrustList:

• 45 open-weight models, 53 closed ones and the licences between them

Sources

• Mistral AI: Mistral Large 4 (Le Chonk), weights release, 29 October 2026
• Hugging Face: Mistral-Large-4-1T-A52B model card and LICENSE, 29 October 2026
• Mistral AI docs: Mistral Large 4 model page and pricing (read 2026-10-30), 30 October 2026
• Mistral AI: Mistral Large 4 public preview, 6 October 2026