Mistral AI Launches Large 4 Preview With 4x the Context Window
The context window is roughly four times that of the previous model. It arrives first as a developer offering, with downloadable model weights planned for the end of October.
Mistral AI released Mistral Large 4 as a public preview on October 6 (local time). The model carries the nickname “Le Chonk.” Starting the day of the announcement, developers can access it through Mistral Studio to wire into their own services. The company said the model weights, the files needed to download and run the model yourself, are scheduled for release at the end of October.
How much more can it read at once?
Feeding in a long document and asking questions about it just got easier. Mistral Large 4 handles both text and images, and its context window, the amount of material it can take in at one time, has grown to 1 million tokens. That is roughly 3.9 times the 256,000 tokens of Mistral Large 3, which was released in December 2025.
The model has grown to about 1.05 trillion parameters, up from 675 billion in the previous version. It uses an architecture that activates only a portion of the model, rather than all of it, when generating a response. Mistral AI said it trained the model on 3,800 Nvidia Grace Blackwell GPUs in its own data centers in Europe.
What do the benchmark scores look like?
Mistral AI claims the model is competitive in coding, security, finance, and locating objects within images. In the company’s own published measurements, the model scored 61.7% on DeepSWE v1.1, 28.3% on Terminal-Bench 4, and 93% on Cybench.
Why are the weights coming only at the end of October?
The release approach differs from the previous model. Mistral Large 3 shipped with its weights available immediately under the Apache 2.0 license. This time, the developer preview comes first. Mistral AI said that before releasing the weights it will run separate testing with security-sector collaborators, vetted partners, and government authorities. Details on the model architecture, additional evaluations, and training methodology are also slated for later disclosure.
What about cost and access?
During the launch discount period, input sent to the AI costs $0.68 per million tokens, and output the AI generates costs $2.09 per million tokens. Reusing stored input when you send the same content again costs $0.07 per million tokens. According to Mistral AI’s official changelog, the 50% discount applies for two weeks.
Pre-discount pricing is $1.36 per million input tokens and $4.18 per million output tokens. Reusing stored input runs $0.14 per million tokens. If you are building a service or plugging the model into business workflows, it is worth budgeting your initial testing costs separately from your operating costs once the discount expires.
For Korean-language users, language coverage is another point of interest. Mistral AI’s language support documentation lists Korean among the supported languages. The company said it plans to roll the service out across multiple regions, and the access path open right now is the developer offering in Mistral Studio.
The standout changes in this release are the expanded context window and the stated plan to release the weights. The next things to watch: whether the end-of-October release lands as announced, and what comes out of the separate security testing and additional evaluations.