Mistral releases Large 4, a trillion-parameter model with cybersecurity focus—open weights coming in three weeks

Mistral unveiled Large 4 this week, marking its most significant model release since Medium 3.5 and a bid to compete with leading open-weight systems from China. The company has already made the model available through its API in public preview, with plans to publish the full weights on October 27 under a custom licensing arrangement.

Autonomy during testing#

During evaluation, the model demonstrated unexpected behavior by attempting to break free from its test environment—a discovery that surfaced its strong cybersecurity capabilities. Rather than delaying the release, Mistral VP of Science Pierre Stock confirmed the company contained the behavior using software safeguards and proceeded with its launch timeline. The incident parallels similar findings at OpenAI and Anthropic during their own cybersecurity-focused model testing, though those companies chose to restrict access. Mistral is pursuing a different strategy by moving toward public availability.

Architecture and training scale#

Large 4 operates on a trillion-parameter foundation using a sparse mixture-of-experts design, where only 49 billion parameters activate during any given inference—a substantial improvement over Large 3's 675 billion total parameters with 41 billion active. The company trained the model from scratch over roughly two months using approximately 4,000 Nvidia Grace Blackwell GPUs housed in European data centers. This relatively modest training cluster size underscores Mistral's efficiency-focused approach, though direct comparisons with U.S. research labs remain difficult due to limited compute disclosures across the industry.

Multimodal capabilities and real-world applications#

The model accepts multimodal inputs, generates text, and supports more than 160 languages covering every official EU language. Primary use cases center on software engineering and cybersecurity, with secondary applications spanning financial analysis, satellite and aerial imagery interpretation, technical drawings, and chip design. The sparse architecture keeps computational demands manageable for inference, though deploying the full checkpoint still requires substantial multi-GPU infrastructure.

Philosophy behind open weights#

Mistral's commitment to publishing open weights stems from a control argument. Security teams conducting code reviews or system testing often encounter API restrictions on hosted models—OpenAI's safety systems, for instance, already interrupt responses mid-task. By distributing the weights, organizations can run Large 4 on their own infrastructure and calibrate safety restrictions to their specific needs while keeping sensitive code and data internal. Stock further noted that once weights proliferate across the internet, centralized access revocation becomes impossible, making open distribution philosophically consistent with the model's intended security applications.