Mistral Large 4 weights are deployable: control and licensing boundaries

Mistral announced the Large 4 preview API on October 6, with weights planned for month’s end.¹ Hugging Face reported their availability on October 29.² The developer’s model repository now lists downloadable files, deployment instructions and a license.³

Archival photograph: October 14, 2011, NERSC data center. Rear of rack at NERSC data center - closeup. Photograph: Chiara Coetzee; CC0 1.0.

Model size is not the cost of each run

The developer's model documentation states that Large 4 uses a mixture-of-experts architecture, with approximately one trillion parameters in total and approximately 49 billion activated per token; including the embedding and output layers, active parameters total approximately 52 billion. The model accepts text and images and outputs text, with a stated context length of 1 million tokens. The developer provides an FP8 checkpoint and a separate NVFP4 checkpoint; the latter targets Blackwell GPUs and is described by the developer as reducing GPU memory requirements.³

These are descriptions of the model's design and supported capabilities, not a guarantee that every organization can process the full context length at low cost. Hardware requirements, GPU memory, throughput and actual expenses still depend on the quantized weight version and runtime configuration. Total parameters, parameters activated per token and context capacity answer different questions; none of these figures can serve as a complete measure of service costs.

Self-deployment also brings governance responsibilities

Open weights change the boundaries of control: within the license's permitted scope, businesses can run the model in their own or chosen environments and make more decisions about data flows, model configuration and availability. Greater autonomy also places responsibility for access management, logging, model updates and safeguards against misuse on the deployer.

The model documentation provides benchmark results for coding and other tasks and describes cybersecurity capabilities. These are developer-reported results, not findings independently reproduced by this site; benchmark scores are not equivalent to detection rates, repair quality or safety guarantees in any given production environment.³ Decisions still require testing on real tasks, distinguishing whether the model provides usable answers, whether the system executes correctly, and whether errors can be detected and contained.

Open weights remain subject to a specific license

The model and related documentation use the Mistral Ecosystem Open-Weight License, a modified MIT license with additional conditions. It cannot be abbreviated to unconditional MIT or Apache open source. The terms permit the relevant use, copying, modification and deployment, while requiring copyright and license notices to be retained in copies or substantial portions.⁴

If a licensee or its affiliate operates a “Model-as-a-Service” business as defined in the terms, a separate commercial license from Mistral must be obtained before using the model or its derivatives for any commercial purpose, except as provided in Section 3. The definition focuses on model access that gives third parties substantial control over inputs, parameters or training data. End products that merely embed a model within a specific function, and services that simply forward requests to other hosted models, fall outside that definition; internal use and access through official Mistral products have separate exceptions. This cannot be generalized into a claim that all hosting or all commercial use is subject to the same restriction.⁴

This release has practical significance for organizations seeking to deploy a large model themselves. Procurement or launch decisions still require checking the license, hardware budget, independent tests and safety governance, rather than reaching a conclusion based on parameter counts or rankings alone. This article edits, synthesizes and translates the model documentation, retaining its attribution notice: 2026 - Mistral; complete original license notice. This does not imply Mistral's endorsement of this article.

Original license notice

Mistral Ecosystem Open-Weight License

Mistral Ecosystem Open-Weight License. This license is a modified version of the MIT license. Attribution notice: 2026 - Mistral

Permission is hereby granted, free of charge, to any person (the “Licensee”) obtaining a copy of this model and associated documentation files - including the model weights, parameters, configuration files, and associated documentation (the “Model”), to deal in the Model without restriction, including without limitation the rights to (a) use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Model; (b) run, deploy, further train or otherwise modify the Model and create derivative works from it; and (c) permit persons to whom the Model is furnished to do so, subject to the following conditions:

1. The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Model.

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. The Licensee may contact Mistral, using this form: https://mistral.ai/contact/, to request a commercial license.

"Model as a Service" means giving a third party access to the Model or its derivative works for inference or post-training by any means (for instance via API or hosted endpoints), in a manner that allows such third party to exercise meaningful control over the inputs, parameters, or training data. This does not include (a) end-user products with model capabilities solely embedded within specific features or harnesses, or (b) mere relaying of requests to models hosted by others.

3. The requirements set forth in Section 2 do not apply to: (a) internal use of the Model or its derivative works, defined as any use that does not make the Model or its derivative works available to third parties; or (b) any use of the Model accessed through Mistral’s official products.

THE MODEL AND ANY OUTPUT THEREFROM ARE PROVIDED “AS IS”, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL MISTRAL BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE MODEL OR THE USE OR OTHER DEALINGS IN THE MODEL.

For any questions regarding this license, please contact us using the form here: https://mistral.ai/contact/

Source notes

1. 2026-10-06 · Introducing Mistral Large 4 · Mistral AI. ↩

2. 2026-10-29 · Welcome Le Chonk: Mistral Large 4 on Hugging Face · Hugging Face. ↩

3. The model documentation does not specify a fixed publication date · Mistral Large 4 - Le Chonk · Mistral AI; this article uses the verified fixed version of the documentation. Parameter counts, capabilities and test results are all developer statements. ↩ ↩² ↩³

4. The license file does not specify a fixed publication date · Mistral Ecosystem Open-Weight License · Mistral AI; the complete original terms govern the specific rights and exceptions. ↩ ↩²