On October 29, 2026, Mistral released downloadable weights for Large 4. Businesses building coding or workplace assistants can now run and adapt the model on infrastructure they control.
Weights are the files that hold a model’s learned parameters. Releasing them gives businesses an alternative to having Mistral process every request. The company’s October 6 preview offered hosted access and promised downloads later in the month; this release delivers those files with an updated model.
Control comes with computing costs
Hugging Face’s published hosting recipe uses eight NVIDIA H200 graphics processors. That is an example configuration, not a universal minimum. It illustrates why downloading the model and operating a workplace service are different commitments.
Capacity also depends on the setup. Mistral’s model card lists a one-million-token context window—the amount of material the model can work with at once, measured in processing units called tokens. But Hugging Face’s example supports about 131,000 tokens and up to 16 simultaneous requests. Buyers handling large documents should ask for the configured capacity, not assume every deployment delivers the headline maximum.
Hosted access remains an option. As of October 29, Mistral’s API rates are $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens. The documentation still labels Large 4 Public Preview. Those usage charges are not a comparison with the full cost of running it yourself.
The licence distinguishes internal tools from model services
The Mistral Ecosystem Open-Weight License permits use, adaptation and distribution subject to conditions, including retaining notices. It is a modified MIT licence, not unrestricted MIT.
Under Section 2, a business whose own operations or affiliates include a model-as-a-service business must obtain a separate commercial licence before commercial use of Large 4 or derivatives. The definition concerns third-party access with meaningful control over inputs, parameters or training data for generating responses or further training.
Section 3 exempts internal use that does not make the model available to third parties, and use through Mistral’s official products. The model-service definition also excludes capabilities solely embedded in specific end-user features or application wrappers, and simple forwarding to another host. Those exclusions do not erase the affiliate condition. A software business needs to consider both its product and its wider business activities. Fees for a separate agreement are not stated in the published licence.
What the release changes for buyers
Mistral reports a 41% score on the Terminal-Bench 4.0 coding benchmark, up 13 percentage points from its preview. That is a reported evaluation result, not a guarantee of correct work or lower operating costs.
Our assessment: hosted access is a practical starting point when the immediate question is whether Large 4 can handle your tasks. Self-hosting becomes a distinct option when deployment control or customization justifies the infrastructure and operating work. Compare quotes for the same workload, including document length, simultaneous users and staffing—not just the availability of downloadable files.
For another model with downloadable weights, see our coverage of Reflection’s Beam release for custom coding assistants. Its licence and hardware guidance are separate from Mistral’s.