Mistral Large 4 Weights Land With a Clause Barring Hosted Inference Without a Licence

Mistral has published the weights for Mistral Large 4, a mixture-of-experts model of about 1 trillion parameters, under a licence it calls the Mistral Ecosystem Open-Weight License. The licence text describes it as "a modified version of the MIT license". Running the model for yourself is free. Selling access to it is not.

Mistral announced the model on October 6 as a public preview through its API and said weights would follow by the end of the month. Artificial Analysis lists the release date as October 29.

What the licence restricts

The grant covers the standard MIT verbs: use, copy, modify, merge, publish, distribute, sublicense and sell copies. It also covers deploying and further training the model, and making derivative works.

The restriction is a Model-as-a-Service clause. A licensee or affiliate that runs the model as a service "must enter into a separate commercial license with Mistral before using the Model". The licence defines the trigger as "giving a third party access to the Model or its derivative works for inference or post-training by any means."

Two carve-outs apply. Internal use that does not make the model available to third parties is exempt, and so is use through Mistral's own products. The licence states no revenue threshold.

The launch post names no licence at all. It says only that the model has open weights and supports self-deployment, so the terms were readable only once the files appeared.

How it compares with Qwen

Alibaba's Qwen team split its August Qwen3.8 release across two licences, eWeek reported. The 27B dense model is Apache 2.0. The 2.4-trillion-parameter Max-class model, with 95 billion active, uses a custom licence that requires a separate agreement from businesses running a model-as-a-service or AI work assistant once aggregate revenue passes $50 million in any consecutive 12 months.

Model	Licence	Hosted-service condition
Mistral Large 4	Mistral Ecosystem Open-Weight License	Separate licence before any third-party inference or post-training access
Qwen3.8-2.4T-A95B	Alibaba Qwen3.8-Max licence	Separate licence above $50 million aggregate revenue
Qwen3.8-27B	Apache 2.0	None

The difference is where the line sits. Alibaba's trigger is a revenue number. Mistral's has no number, so a two-person company reselling a fine-tune through an API falls inside it. For earlier coverage of a Chinese lab's licence terms, see MiniMax M3.1 Flash. For a permissive counter-example, see Recursive's Apache-licensed Flint.

Size, price and scores

The card gives 49B active parameters in one place and about 52B in another, which it says includes embeddings and output layers. Artificial Analysis lists 1.1 trillion total and 52 billion active. The architecture is 128 routed experts with 4 active per token plus one shared expert. Context is 1M tokens, input is text and images, and output is text.

Mistral's API charges $0.70 per million input tokens and $2.10 per million output tokens, per Artificial Analysis. That page scores the model 26 on its Intelligence Index, against a median of 23 for open-weight models it compares, and Mistral cites the same figure. Artificial Analysis computed it, so it is the independent number.

Most other benchmark figures in the launch post are Mistral's own, and the post does not name the evaluator for Terminal-Bench 4.0 (28.3%) or Cybench (93% solved). One third-party result it cites is Surge AI's blind human coding evaluation, where Mistral Large 4 ranked second of five at 3.74, behind Claude Opus 5 at 4.22.

Running it

The weights are downloadable with no gating form. The card says Mistral trained on 3,800 NVIDIA Grace Blackwell GPUs and recommends 8-GPU tensor parallelism with vLLM or SGLang. An NVFP4 checkpoint targets Blackwell, and GGUF builds exist for llama.cpp. The card says no Hugging Face Inference Provider deploys it yet.

The card gives no release date, and the licence carries an attribution notice dated 2026. Neither the licence text nor the card gives a price or terms for the commercial agreement.