Mistral makes Large 4 AI model available for download
The French developer claims gains in coding and cyber defence, while allowing organisations to run the model without relying exclusively on a hosted service.
French AI developer Mistral has released downloadable weights for Mistral Large 4 (ML4), its latest flagship model. The company claims improvements in coding and says the model gives businesses and governments cyber-defence capabilities for detecting, analysing and responding to threats affecting code and IT infrastructure.
Organisations can use the weights to operate ML4 on infrastructure they control, potentially keeping sensitive information such as proprietary software code away from external AI services. They still need appropriate hardware and expertise to run it securely. Mistral says local operation also avoids the risk of another provider limiting or disabling access.
After presenting a preview on 6 October, Mistral continued reinforcement learning and conducted real-world testing with customers and partners. It added two reasoning modes; users now have low, medium and high settings to match different tasks.
Mistral puts average token use per task at 47 per cent below that of the Chinese GLM-5.3 model. It also says additional training increased ML4’s score from 28 per cent to 41 per cent on Terminal-Bench 4.0, a test of coding tasks performed by AI agents in a terminal.
According to the company, ML4 has a total of 1.05 trillion parameters and activates around 52 billion. Training took place at Mistral’s European data centres using 3,800 Nvidia Grace Blackwell GPUs. Access is available via Mistral’s API, with downloadable weights hosted on Hugging Face.
