Mistral announced the release of Mistral Large 4’s downloadable weights on October 29, 2026, moving the model from its public-preview stage toward self-hosting. The preview was listed on October 6; NeoTeo covered it in its report on the public preview. The weights are available under a license with a specific condition for some commercial deployments.
What the released checkpoint supports
The Hugging Face checkpoint is named mistralai/Mistral-Large-4-1T-A52B; Mistral’s API model name is mistral-large-4. The model accepts text and images and returns text, with a stated context window of 1 million tokens. Mistral describes it as a multimodal Mixture-of-Experts model with about 1.05 trillion total parameters. Its model card counts 49 billion active parameters per token when embeddings and output layers are excluded, or 52 billion when they are included. Mistral’s model card
Amine AI’s October 30 video includes a creator-made prompt comparison between Mistral Large 4 and Claude Opus 5.5, alongside screens showing model cards and benchmark graphs.
What Mistral’s vLLM example specifies
Mistral’s published vLLM example sets --tensor-parallel-size 8. Tensor parallelism splits model computation across processors; this setting describes the example, not a stated minimum GPU count. If GPU memory is insufficient, Mistral advises lowering --max-model-len, which controls the maximum input length, and/or --max-num-seqs, which controls how many sequences are processed concurrently. Mistral’s vLLM example
That makes context length and concurrency the first configuration variables to consider when the example runs short on memory. Mistral also lists SGLang, Transformers, and llama.cpp as supported deployment software, and recommends vLLM for production inference.
When the license calls for a separate commercial agreement
The repository uses the Mistral Ecosystem Open-Weight License, which describes itself as a modified MIT license. It permits uses including modification, distribution, sale, deployment, and further training, subject to its conditions. One condition concerns a licensee or affiliate that operates a qualifying Model-as-a-Service (MaaS) business. Read the license
The license defines MaaS access as giving a third party access for inference or post-training with meaningful control over inputs, parameters, or training data. Here is how its stated terms distinguish that arrangement from several others:
| Deployment arrangement | License treatment |
| Internal use without making the model or derivatives available to third parties | Exempt under the license’s internal-use provision. |
| A licensee or affiliate operates a qualifying MaaS business | A separate commercial license with Mistral is required before commercial use of the model or derivatives. |
| Model capabilities are embedded solely in specific end-user product features or harnesses | Excluded from the license’s MaaS definition. |
| Requests are merely relayed to models hosted by others | Excluded from the license’s MaaS definition. |
API rates and benchmark task cost measure different things
Mistral lists API rates of $0.70 per million input tokens, $0.07 per million cached input tokens, and $2.10 per million output tokens. Those are token-based API rates. Separately, Artificial Analysis lists $2.06 per Intelligence Index task for Mistral Large 4 (high), alongside an index score of 26 and a 1-million-token context window. That task-cost figure is a benchmark measure, not an estimate of what a particular application will cost. Mistral’s model documentation · Artificial Analysis leaderboard
