French AI lab Mistral has released Mistral Large 4 (ML4), nicknamed "Le Chonk" for its trillion-parameter size, now in public preview. The open-weight model is positioned as a cybersecurity tool that enterprises and governments can fully own and control, avoiding the risk that a closed-model vendor could revoke access. Mistral says this addresses a key concern for clients after a summer of AI security incidents.
Mistral co-founder Guillaume Lample said the model's cyber defense capabilities let organizations defend against threat actors who jailbreak closed models. The company claims ML4 outperforms models from Kimi, DeepSeek, and Meta on cyber capabilities. Early third-party benchmarks also show it competing with pricier proprietary models on some computer-vision tasks and setting a record for open-weight models on Harvey's Legal Agent benchmark, though final benchmarks are still pending and scores should be treated cautiously.
Training efficiency is a notable contrast: Mistral trained ML4 from scratch on 4,000 Nvidia Grace Blackwell GPUs over two months in its European data centers, while Nvidia CEO Jensen Huang said OpenAI used roughly 100,000 GPUs for GPT-6 Astra. Mistral will release the model weights on Oct. 27 after a month of testing with developers and security authorities. The company also emphasized sovereignty, promising flexible deployment options including a European region where data stays under EU jurisdiction.