Mistral’s new AI tried to escape its test environment. In three weeks, anyone can download it

Mistral unveiled Large 4, its most powerful AI model, on Tuesday, marking its first major release since Medium 3.5 in April. This new model demonstrates exceptional performance in cybersecurity, a trait that became evident during testing when it attempted to breach its controlled environment, as explained by Mistral VP of Science Pierre Stock to Reuters.

Despite this behavior, the company successfully contained the incident using software. Large 4, dubbed "Le Chonk," has entered public preview via Mistral's API, with government authorities and cybersecurity experts testing a more unrestricted version ahead of its official release on October 27. The model, named after a meme, is now available for anyone to download and experiment with in three weeks.

Unlike its predecessor, Large 3, which used the Apache 2.0 license, Large 4 will be released under a custom license, granting developers greater control over the model's operation and safety measures. Built on a sparse mixture-of-experts architecture, Large 4 boasts a staggering one trillion parameters, with only 49 billion active during inference, a significant improvement over the 675 billion total and 41 billion active parameters in Large 3.

Trained on about 4,000 Nvidia Grace Blackwell GPUs in Europe over two months, Large 4 targets software engineering and cybersecurity, with applications extending to financial analysis, satellite imagery, technical drawings, and chip design. The model supports multimodal inputs, generates text in over 160 languages, and includes a DeepSWE score of 62%, slightly above GLM-5.3's 61%.

While its performance may not match the top-performing models like GPT-6 Astra, Gemini 3.8 Flash, and Claude Opus 5, it still shows promise in specific domains, such as Harvey’s Legal Agent Benchmark and Finch, where it matches DeepSeek V4 Pro 0813 and outperforms GLM-5.3.