On October 29, 2026, European artificial intelligence laboratory Mistral AI officially released the weights for its flagship model, Mistral Large 4. Known colloquially by the internal nickname Le Chonk, the model is now available for download following an initial preview phase earlier in the month. Distribution is governed by the Mistral Ecosystem Open-Weight License, granting developers and researchers direct access to the model parameters. This release marks a significant step for the company in providing frontier-class weights to the broader technical community.
At the core of Mistral Large 4 is a massive Mixture-of-Experts architecture containing approximately one trillion parameters in total. Despite this vast scale of roughly 1,000 billion parameters, the system dynamically routes computation so that only 49 to 52 billion parameters are active per token. This sparse activation mechanism ensures high operational efficiency during inference, avoiding the severe computational penalties typical of dense architectures. Consequently, the model combines enormous representational capacity with viable execution speeds.
The model is equipped with a context window spanning up to one million tokens, allowing it to ingest and analyze extensive codebases and lengthy documents in a single prompt. Furthermore, the architecture provides native multimodality across both text and image modalities. Visual input is processed directly through an integrated Pixtral vision encoder, eliminating the need for detached vision frontends. This unified design enables coherent visual comprehension alongside deep contextual reasoning.
To train the trillion-parameter system, Mistral AI relied on a high-density cluster situated within European data centers. The infrastructure comprised 3,800 NVIDIA Grace Blackwell GPUs dedicated to managing the complex training runs. By conducting the entire compute run on European soil, the laboratory demonstrated that frontier-scale foundation models can be produced on regional infrastructure. This operational choice reflects expanding European capabilities in high-performance computing for machine learning.
Initial benchmark evaluations indicate that Mistral Large 4 delivers competitive performance across demanding technical domains. In evaluations dedicated to agentic coding, specifically Terminal-Bench and FrontierCode, the model achieved new record scores among open-weight systems. It also demonstrated high capability in defensive cybersecurity workflows, recording an 82 percent score on the Artificial Analysis Cyber Index. These results show that open-weight architectures are increasingly closing the performance gap with proprietary frontier models in specialized software tasks.
By distributing the model under the Mistral Ecosystem Open-Weight License, the organization provides enterprises and researchers with an inspectable foundation for complex workloads. The blend of agentic programming aptitude, native visual understanding, and defensive cyber capabilities allows for diverse production deployments. Engineering teams can now host and adapt the MoE model locally without relying on closed cloud endpoints. As a result, the release establishes a new technical baseline for what open-weight models can achieve in enterprise environments.