Mistral Large 4: a trillion-parameter multimodal model from Europe
Mistral released a new flagship in public preview at $0.7 / $2.1 per million tokens. Open weights are promised by the end of October.
Mistral has released a public preview of Mistral Large 4, nicknamed le Chonk. It’s a large open-weight model trained in Mistral’s own European datacenters.
What’s new
- Size: about 1T total parameters and 52B active, using Mixture-of-Experts.
- Multimodal: the model takes more than text natively.
- Languages: training data spans over 160 languages, including all official EU languages.
- Training: from scratch on 3,800 NVIDIA Grace Blackwell GPUs.
According to Mistral, it scored 61.7% on DeepSWE v1.1 and 82% on a reproduce-and-patch vulnerability test. In a human coding evaluation by Surge AI, it placed second of five, behind only Claude Opus 5.
How to try it
The model is available through the API in Mistral Studio at $0.7 per million input tokens and $2.1 per million output tokens. It can run on Mistral’s European infrastructure, in a private cloud or on premises.
Limitations
This is a preview, and the weights are only promised by the end of October. The announcement page doesn’t name a license, so check commercial terms separately. Most of the numbers come from the company itself.
Prepared with AI from public sources.