Weights released at month's end as promised
France's Mistral AI on Oct. 29 released on Hugging Face the weights of an improved version of "Mistral Large 4" (nicknamed Le Chonk), a large model with 1 trillion parameters. When it released the preview on Oct. 6, the company had said it would release the weights at the end of the month.
The model uses a mixture-of-experts architecture that activates only some of its expert parts, using about 52 billion parameters per token. It can also read images and handles 1 million tokens at a time. It was trained from scratch on 3,800 chips from NVIDIA. The company lists delegating programming tasks and defending against cyberattacks as its main uses.
Major improvement over the preview
In the improved version, the score on the terminal operation evaluation Terminal-Bench 4.0 rose from the preview's 28.3% to 41%, which the company says ranks first among models with publicly released weights. The score on an evaluation of resistance to adversarial instructions rose from 93.3% to 97.1%. On "FrontierCode 2.0," a software development evaluation by Cognition, it scored 19.7%, which the company says places it near the top among open models.
| Evaluation | Preview | Improved version |
|---|---|---|
| Terminal-Bench 4.0 | 28.3 | 41 |
| Lakera B3 (resistance to attacks) | 93.3 | 97.1 |
Falls short of Chinese open models in some respects
It scored 26 points on Artificial Analysis' overall index (version 5), and Mistral itself acknowledges in the model description that top Chinese open models score higher. Eleven organizations, including CrowdStrike, Cloudflare and Cisco, provided endorsements of its cyber defense performance. Customers can customize it with "Mistral Forge," which allows tuning on their own security data.
API pricing is $0.70 per million input tokens and $2.10 per million output tokens. The license is a modified MIT license, and using the model in a business that provides AI to other companies requires a separate agreement with Mistral.
