Mistral AI Releases Mistral Large 4, and You Can Download and Run It Yourself

Mistral AI released the weights for Mistral Large 4 on October 29 (local time). Weights are the files that hold what a model has learned. Users can download them from the company’s official Hugging Face repository and run the model in their own environment.

What changes when you run it yourself

Users now get to choose where the model runs. On October 6, Mistral AI opened a public preview through its API, the interface developers use to plug a model into their own services. The company said at the time that the weights would follow by the end of the month, and that has now happened, so there is a second path alongside calling Mistral AI’s API: installing and running the model yourself.

Mistral Large 4 takes text and images as input and responds in text. Internally it is built with multiple specialized sections, activating only the ones it needs. It can read up to 1 million tokens at a time, a token being a small chunk that text is broken into. It can also call external tools, and the amount of effort it puts into producing an answer can be adjusted.

Coding performance against rival models

Mistral AI says the model delivers comparable coding performance while using fewer tokens. The company put its score on the Artificial Analysis Coding Agent Index at 48%, roughly level with GLM-5.3. The average number of tokens spent per task was 47% lower than GLM-5.3. For users considering the model for coding, the figures provide a basis for comparing performance and token use.

Other benchmarks were less favorable. The Intelligence Index v5 score that Mistral AI lists in the model card is 26 points. The company notes that this falls below leading open models from China.

Cost and conditions

The API remains available as before. The pricing listed in the official documentation is $0.70 per million input tokens and $2.10 per million output tokens. Reusing cached input costs $0.07 per million tokens. Users who run the model themselves trade those fees for the cost of operating their own hardware.

Conditions vary depending on the business model. Both commercial and non-commercial use are permitted. However, running a “model-as-a-service business” on top of this model requires a separate commercial license from Mistral AI.

In the model card, Mistral AI says the model supports more than 160 languages. Anyone considering deployment in Korea will need to separately verify how well it handles Korean-language documents and queries, as well as what hardware and costs self-hosting would entail.

With this release, Mistral Large 4 moves from a model you call through an API to one you can install and run yourself. Companies and developers can now decide on their own execution environment. Whether to adopt it comes down to performance on the tasks you intend to assign, operating costs, and the licensing terms that apply to your particular business model.