The Paris-based company is publishing the weights of its model with around one trillion parameters. Companies can now run it on their own hardware and adapt it to their own data.
The Paris-based AI company Mistral has published the weights of its language model Mistral Large 4, nicknamed Le Chonk. In October there was initially only a preview version that could be used through Mistral's interface. Now the model can be downloaded from Hugging Face and run on your own servers.
Large 4 has around one trillion parameters and processes images as well as text. According to Mistral, it was trained from scratch in Europe. The licence is a modified MIT licence. Using, modifying and deploying it commercially is free, only those who offer the model as a service to others need a separate licence from Mistral.
Retrained for programming work
Since the preview, Mistral has spent three weeks retraining the model with reinforcement learning. According to the manufacturer, its score on the Terminal-Bench 4.0 test rose by 13 percentage points to 41 percent, while the model outputs 30 percent fewer tokens. Mistral's second focus is defence against cyberattacks. The European cybersecurity agency ENISA says it is testing Large 4 to find security vulnerabilities and audit program code.
Through Mistral's interface, usage costs 0.70 dollars per million tokens for input and 2.10 dollars for output. Large 4 is also offered on Microsoft Foundry and through NVIDIA's computing services.
This is interesting for companies that do not want to hand their data to a cloud provider. A model of this size does, however, need a server with many powerful graphics processors.