Mistral unveiled Mistral Large 4 on Tuesday, a 1-trillion-parameter open-weight model trained over two months on 4,000 Nvidia Grace Blackwell GPUs deployed across European data centers. The company claims it is the strongest open model developed outside China.
ML4, also known as le Chonk, is currently available in preview for developers, cybersecurity leaders and government authorities, with broader release planned later this month. The model performs specialized tasks in cybersecurity, coding, manufacturing, finance and multimodal applications, though it still lags frontier models in coding benchmarks.
The effort reflects Mistral's capital-intensive strategy to close the gap with Chinese competitors like Deepseek, which have gained substantial adoption worldwide. Mistral secured a 3 billion euros ($3.4 billion) Series D in September at a 21 billion euro valuation, giving it one of the largest war chests among European AI startups.
Guilaume Lample, Mistral's co-founder and chief scientist, said the company plans to expand training capacity using the fresh capital. He highlighted the model's cyber defense capabilities, positioning it as a tool for enterprises and governments to counter threat actors who exploit closed systems.
Open-weight models, which can be self-hosted and modified, contrast with closed systems from OpenAI and Anthropic. Their appeal lies partly in avoiding proprietary guardrails. Earlier this year, Hugging Face used China-based GLM 5.2 to mitigate a security breach after leading U.S. closed systems misidentified its defensive actions as attacks.
