Mistral lets you download its biggest AI model and run it on your own servers

Mistral has released the downloadable files for Mistral Large 4, its biggest AI model, giving businesses the option to run it on their own infrastructure instead of sending requests through a hosted service.

The October 29 release follows an October 6 preview. The model accepts text and images and produces text, with room for up to one million tokens of context, the material it can consider in a request.

You can modify and fine-tune it for commercial work. But the license draws a boundary around businesses that sell access to models, so downloading the files does not give every provider the same commercial rights.

Running the model yourself comes with real hardware demands

The official model card lists one trillion parameters, the numerical values learned during training. Only part of the model works on each token, but its full size still makes deployment a substantial infrastructure project.

Mistral’s example setup spreads the model across eight graphics processors. That is a sample configuration, not a promise that any eight GPUs will be sufficient.

Hugging Face’s release guide lists support across deployment tools, including vLLM, Transformers and llama.cpp. It also describes compressed versions that reduce memory demands.

If your team wants to keep working through a hosted service, Mistral still offers its API. The announced rates are $0.70 per million input tokens, $0.07 for cached input and $2.10 per million output tokens. Those charges do not describe the cost of running your own servers.

The license separates internal use from selling model access

The Mistral Ecosystem Open-Weight License requires a separate agreement for commercial use by businesses, or their affiliates, operating a Model-as-a-Service offering.

It defines that as giving third parties access with meaningful control over inputs, settings or training data. Products that embed the model within specific features are excluded from that definition.

The extra licensing requirement also does not apply to internal use that keeps the model unavailable to third parties, or to access through Mistral’s official products.

For an in-house assistant, the release opens a route to controlling where the model runs. For a business selling flexible access to it, the license needs a closer look before launch.

Source: Mistral AI