Mistral released the weights for Mistral Large 4 on 29 October 2026, after introducing the model in public preview earlier in the month. The release gives teams a downloadable checkpoint and a route to running the model on their own infrastructure, alongside Mistral’s hosted API.
What is in the checkpoint?
Mistral Large 4 is a mixture-of-experts model with roughly 1 trillion parameters in total. Its model card says about 52 billion parameters are active per token when embeddings and output layers are included. It accepts text and image inputs, produces text, and has an advertised context window of 1 million tokens. The same weights support low, medium and high reasoning settings.
The active-parameter figure describes how the model processes a token. The much larger total weight count remains relevant when planning storage and serving capacity. A stated context limit also does not establish the cost or quality of a particular long-context workload.
How can teams run it?
The model card provides serving instructions for vLLM and SGLang. Mistral also offers the model through its API as mistral-large-4, according to its model documentation. Teams evaluating self-hosting will need to test the released checkpoint against their own hardware, latency, throughput and operational requirements.
What does the licence allow?
The weights carry the Mistral Ecosystem Open-Weight License. It permits use, modification and distribution subject to its conditions, including retaining the licence notice. It also requires a separate commercial agreement before a business operating Model-as-a-Service uses the model or its derivatives commercially.
The licence defines that service category in terms of third-party access with meaningful control over inputs, parameters or training data. It sets out exceptions, including internal use and certain products with model capabilities embedded in specific features. A team offering hosted access should check the actual licence against its planned service.
What remains to be tested?
Mistral reports improved coding and defensive-security benchmark results since the preview. Those results are claims in the company’s release; they do not establish how the downloadable checkpoint will perform in a reader’s environment. Teams should test the tasks, serving configuration and reasoning settings they intend to use.
