Agentic data science with the Data Agent Kit
cloud.google.com·6h ago
Data Science
Simon Willison's Blog·6d ago·5 min read
Introducing Mistral Large 4: Le chonk
TL;DR
Mistral has unveiled Mistral Large 4, a 1 trillion parameter model trained on 3,800 NVIDIA GPUs, showing significant improvements over its predecessor. The model is available via API, with open weights expected soon.
✦ Why It Matters
Explore Mistral Large 4 via the API to evaluate its capabilities for your projects.
Key Takeaways
- Test Mistral Large 4 through the API for your applications.
- Monitor the release of open weights for further experimentation.
- Compare performance metrics with other models like DeepSeek 4.1 Flash.
- Evaluate reasoning levels to optimize output quality.
Full Summary
Mistral Large 4, dubbed 'Le chonk', is a new large language model featuring 1 trillion parameters and 49 billion active parameters, trained on a cluster of 3,800 NVIDIA Grace Blackwell GPUs. This model is currently accessible through the Mistral API, with plans to release its open weights by the end of the month.
It supports two reasoning levels: 'none' and 'high', with the latter producing better outputs despite using fewer tokens. In terms of performance, it scored 38 on Artificial Analysis, significantly improving from the previous Mistral Large 3, which scored only 9.
While it doesn't reach the capabilities of top-tier models like Fable, it positions Mistral as a competitive player, approximately six months behind the leading edge in large language models.