Mistral AI, the French artificial intelligence laboratory, announced the public preview of Mistral Large 4 (ML4), a multimodal model with 1 trillion parameters, internally codenamed "Le Chonk." The model, which has 49 billion active parameters during inference, was trained from scratch over approximately two months using 4,000 Nvidia Grace Blackwell GPUs in Mistral's European data centers. Mistral states that ML4 was trained across more than 160 languages, including all official languages of the European Union.
The company has made ML4 available through its API for public preview, with plans to publish the model's weights on October 27. This three-week testing period involves developers, cybersecurity leaders, and government authorities. Mistral intends for the open-weight release to allow enterprises and governments to run and customize the model on their own infrastructure, including with zero-data-retention options. Guillaume Lample, Mistral co-founder and chief scientist, stated that ML4 is at the forefront of open-weight models and that its capabilities are expected to improve as reinforcement learning concludes and training capacity expands.
Mistral is emphasizing ML4's performance in specialized enterprise applications. The company reports that ML4 achieved a 62% score on DeepSWE v1.1, a software engineering benchmark. It also claims a 67% score on Finch, an enterprise finance and accounting workflow benchmark. In cybersecurity, Mistral reports ML4 scored 93% on Cybench and 82% on CyberGym-E2E, a test that requires models to reproduce and patch vulnerabilities. Mistral highlights that some closed frontier models score near zero on CyberGym-E2E because they refuse the task.
The development of ML4 reflects Mistral's strategy of providing efficient, openly available model weights. The company has expanded its offerings beyond model weights to include developer and enterprise products, customization services, and inference infrastructure. This broader effort aims to provide enterprises with control over the infrastructure, deployment, and engineering surrounding their AI models. Mistral's pricing for ML4's API access is $1.36 per million input tokens and $4.18 per million output tokens.
The release of ML4 follows a viral internet meme earlier this year about a fictional large Mistral model called "Le Chaton Fat." Mistral executives have acknowledged the meme, with Lample suggesting ML4 could be seen as an initial version of the idea, with larger models to follow. The company stated the "Le Chonk" nickname deliberately references the community that supported Mistral building a massive frontier model.