Mistral announced the availability of downloadable Mistral Large 4 weights on October 29, 2026, after opening a public preview on October 6. The release lets organizations inspect, customize, and run the model on infrastructure they control, while its license sets a specific condition for some commercial services. Mistral’s announcement follows NeoTeo’s coverage of the public preview.
When the license requires a separate commercial license
The Mistral Ecosystem Open-Weight License permits commercial use, modification, distribution, deployment, further training, and derivative works, subject to its terms and required notices. A separate commercial license is required before commercial use if the licensee or an affiliate operates a defined Model-as-a-Service business. The license defines that service as providing third parties with inference or post-training access when they can meaningfully control inputs, model parameters, or training data. Access through Mistral’s official products, internal use that does not expose the model or derivatives to third parties, requests relayed to models hosted by others, and end-user products with model capabilities confined to specific features or harnesses are among the stated exceptions to those requirements.
Why the model card gives both 49B and 52B
The figures describe different parameter counts. Mistral Large 4 has 1 trillion total parameters; 49 billion are active per token, while the 52 billion figure includes embeddings and output layers. The model card describes text and image inputs, text output, and a context window of up to one million tokens. Hugging Face’s model card also lists low, medium, and high reasoning effort, with high as the default.
The official serving example uses eight-way tensor parallelism
The model card’s vLLM example sets tensor parallelism to eight. That is a serving configuration, not a universal minimum GPU count. If GPU memory is insufficient, the card advises reducing the maximum model length or the number of concurrent sequences. The settings that fit a deployment depend on its available memory, context length, and concurrency.
Amine AI’s English-language video shows one same-prompt coding task: building a 3D satellite-map visualization with Mistral Large 4 and Claude Opus 5.5. It is the creator’s single-run demonstration, not a standardized benchmark.
