Mistral releases Large 4's weights, but anyone selling it as an API now needs Mistral's permission

Mistral has published the weights of Mistral Large 4, its 1-trillion-parameter flagship, on Hugging Face. The model went into public preview earlier this month; the October 29 release is the improved version after three more weeks of reinforcement learning. It is the largest model Mistral has ever released openly, and it comes with a licence that is noticeably less open than the last one.

The licence

Mistral Large 3 was Apache 2.0, and Mistral's own models page still lists it that way. Large 4 ships under a new Mistral Ecosystem Open-Weight License, described as "a modified version of the MIT license." The MIT part grants the usual rights: use, modify, fine-tune, redistribute, sell. The modification is clause 2:

If the Licensee or any of its affiliates operates a Model as a Service business, the Licensee must enter into a separate commercial license with Mistral before using the Model or its derivative works for any commercial purpose.

"Model as a Service" means giving a third party access to the model or a derivative "for inference or post-training by any means (for instance via API or hosted endpoints)" in a way that lets them exercise "meaningful control over the inputs, parameters, or training data." Two things are carved out: end-user products where the model is embedded in specific features or harnesses, and simply relaying requests to a model someone else hosts. Internal use, where nothing is exposed to third parties, is exempt altogether.

In practice: a company that runs Large 4 on its own servers for its employees, or builds it into a product feature, can do so freely. An inference provider that wants to list it next to GLM or DeepSeek has to negotiate with Mistral first, and so do derivatives, so a fine-tune doesn't get around it. Note also the scope: the condition attaches to any licensee whose affiliates run a MaaS business, and then covers "any commercial purpose," not only the hosted endpoint. A large cloud company that wants Large 4 in a commercial product would seem to need the separate licence even for uses that aren't themselves an API. That is our reading of the text; Mistral hasn't published guidance on it.

On Hacker News, commenters quickly pointed out that this fails the OSI's Open Source AI Definition, which requires use for any purpose without asking permission. Mistral doesn't claim otherwise: it calls the release "open-weight", not open source. The more useful observation is commercial. The one group the licence fences off is the one that would compete with Mistral's own API on price.

The model

According to the model card, Large 4 is a mixture-of-experts model with about 1T total parameters, 49B active per token (52B counting embeddings and output layers), 128 routed experts with 4 active plus one shared expert, a 1M-token context and image input. It was trained from scratch on 3,800 Grace Blackwell GPUs in Mistral's own European datacenters. The checkpoint includes a built-in speculative-decoding draft head, and an NVFP4 quantization for Blackwell GPUs is published alongside. The release also brings three reasoning efforts (low, medium, high), with low replacing the preview's no-reasoning mode.

Mistral's announcement leans on coding and security. Its claims include 19.7% on Cognition's private FrontierCode 2.0, ahead of the open models it compares against, 41% on Terminal-Bench 4.0 (up 13 points from the preview, with 30% fewer output tokens), and 48% on the Artificial Analysis Coding Agent Index, which Mistral describes as tied with GLM-5.3 while using 47% fewer tokens per task. Several of the cyber and coding charts are labelled as Artificial Analysis results; the rest are Mistral's own or its partners'.

Artificial Analysis' own model page is more sober about general capability. It gives Large 4 (high) 26 on its Intelligence Index, against a median of 23 for comparable open models, and calls it fast (161 output tokens per second) but "somewhat verbose": 110M tokens to run the index, versus a median of 100M. API pricing is $0.70 per million input tokens and $2.10 per million output tokens, about half of what the preview cost according to users comparing the two.

A notable admission

The most interesting paragraph in the announcement is about cyber safety. Mistral says Large 4 refuses 96.6% of malicious cyber requests on average across the cyber prompts of JailbreakBench, StrongREJECT and AgentHarm, the highest it measured among open models. Then it says that while monitoring cyber-related use of the preview in its live traffic, "we see a potential for uplift to conduct real attacks," and that it is "modifying our guardrails accordingly."

That is an unusual thing to say in the same post that releases the weights. Guardrails Mistral adjusts on its API don't follow the weights to someone else's GPUs; the only safeguards that travel are the ones trained into the model. Mistral is also giving vetted enterprise partners, such as Orange Cyberdefense, a more capable cyber-specialised variant for penetration testing and red-teaming, which is not part of the open release.