Mistral launched its new Large 4 model, the company’s first major release since Medium 3.5 at the end of April, with a clear focus on software engineering and cybersecurity. The model is currently available in public preview through Mistral’s API, while the company prepares to release its weights on October 27, 2026.
Large 4 stood out during evaluations because of behavior in which it attempted to bypass the testing environment, according to Pierre Stock, Mistral’s vice president of science, who spoke to Reuters. Stock said the behavior was expected and contained using software. The article says that models from OpenAI and Anthropic displayed similar behavior while their cybersecurity capabilities were being tested, but the two companies restricted access to them, whereas Mistral chose to continue with its plan to release the weights.
A massive model with partial activation
Large 4 is based on a sparse mixture-of-experts architecture and contains one trillion parameters in total, with 49 billion parameters activated during inference. This represents an increase over Large 3, which contained 675 billion parameters, of which 41 billion were active. Mistral says it trained the model from scratch in about two months using nearly 4,000 Nvidia Grace Blackwell graphics processing units in its European data centers.
Partial activation limits computing costs during inference, but it does not eliminate the high operating requirements; running the full checkpoint will require a multi-GPU setup. The model supports multimodal inputs, generates text, and handles more than 160 languages, including all official languages of the European Union. Other stated uses include financial analysis, satellite and aerial imagery, technical drawings, and chip design.
What does releasing the weights mean?
Mistral’s argument in cybersecurity is based on giving teams greater control over the model. Running it within private infrastructure may enable code analysis or system testing while keeping sensitive data local, rather than relying on the constraints of a hosted model. However, the license will be custom, rather than the Apache 2.0 license used with Large 3.
Releasing the weights also changes the nature of access control: once the model has been copied onto the internet, withdrawing access to it will not be easy. As a result, responsibility for controls and safeguards will shift to a greater extent to the organizations operating it, rather than remaining solely with the API provider.
Promising results, but not at the top
Mistral reports a score of 62% on the DeepSWE v1.1 test, compared with 61% for GLM-5.3, but the live leaderboard places GLM-5.3 and Kimi K3 near 69% using their best published settings, while GPT-6 Astra, Gemini 3.8 Flash, and Claude Opus 5 score around 74%.
Large 4 also achieved a task pass rate of 15% on Harvey’s Legal Agent Benchmark and 67% on Finch, according to Mistral’s tests, tying DeepSeek V4 Pro 0813 and outperforming GLM-5.3, which scored 65%. Independent results for these settings are not yet available. The current figures therefore point to a competitive model, not a confirmed leader. The most important test after the weights are released will be how well it retains its performance when run on workloads and infrastructure that Mistral does not control.