Mistral lets companies host its biggest AI, Large 4, but not sell access to it

Three weeks after a preview reserved for partners, Mistral AI released the final version of Mistral Large 4, nicknamed “Le Chonk,” on October 29. Its weights, meaning the parameters of the trained model, can be downloaded freely from Hugging Face. A company can therefore run it on its own servers, without sending its data to an outside service.

There is one catch, written into the license. Mistral publishes it under a “Mistral Ecosystem Open-Weight License,” a modified MIT license. It allows use, modification, retraining and commercial use. There is a single exception: a company that offers the model to others as an online service (“Model-as-a-Service”) must first obtain a separate commercial license from Mistral.

What the license changes for a company

For an IT department that wants an internal assistant, a coding agent or a security analysis tool hosted in-house, the license is no obstacle. It targets hosting providers and platforms that would resell access to the model. It is a way for Mistral to open up its model while protecting its API business from cloud providers. It is therefore not a free license in the strict sense, like Apache 2.0.

The model remains heavy. Mistral states 1,000 billion parameters in total, with a “mixture of experts” architecture that activates only some of them for each request. The repository name on Hugging Face indicates 52 billion active parameters, whereas the preview announced 49. At this size, you need clusters of data center GPUs. The model is also available through Mistral's API, OpenRouter, NVIDIA's services and Microsoft Foundry, which Microsoft presents as the first cloud to offer it.

The API price has dropped significantly: $0.70 per million input tokens and $2.10 for output, according to Mistral. For the preview, Next cited $1.36 and $4.18.

Scores mostly published by Mistral

Mistral attributes its progress to several additional weeks of training. On Terminal-Bench 4.0, a test of command-line tasks, it reports 41%, versus 28% for the preview, with 30% fewer tokens generated. The text does not say who measured this score. Other figures come from third parties, but they are presented by Mistral itself: 48% on Artificial Analysis's coding agent index, tied with China's GLM-5.3, and a preference from the Surge Code platform over Kimi K3 and DeepSeek V4.1 Flash, behind Anthropic's Claude Opus 5.5.

The company also highlights cybersecurity, as it did with the preview announcement. It cites a 96.6% refusal rate for malicious requests, measured by Mistral itself on three public test sets. Releasing openly a model presented as strong in cybersecurity raises an obvious question: its safeguards can be removed through retraining once the weights are downloaded, as with any open model.

What remains to be verified

Independent evaluations of the final version are still partial. Artificial Analysis found the preview very verbose, which increases the real cost per task. The coming weeks will show whether the final version fixes this flaw, and whether European hosting providers apply for the commercial license to offer Large 4 to their customers.