Mistral Large 4

An open language model from Mistral AI with about a trillion parameters, about 52 billion of them active at each step. It reads text and images and answers in text, with a context window of one million tokens. Reasoning depth is set per request: low, medium or high. It supports function calls and answers that follow a strict JSON schema. The authors claim fluency in more than 160 languages and single out coding, cybersecurity, finance and law.

The authors recommend running it with vLLM, and SGLang, Transformers and llama.cpp are supported too. The files take up more than a terabyte, and the vLLM example is set for eight GPUs; the card gives no exact memory requirements. A compressed NVFP4 version targets Blackwell GPUs. The model is also available through the Mistral API.

The weights are public without any acceptance step. The Mistral Ecosystem Open-Weight License is based on MIT: you may use, change, fine-tune and sell the model if you keep the license text. A company that runs a Model as a Service business, itself or through affiliates, needs a separate commercial license from Mistral at any revenue. Model as a Service here means giving third parties access to the model through an API or hosted service with control over inputs, parameters or training data. Internal use and finished products with the model embedded in specific features are outside this condition. That is why we mark commercial use as limited.