Mistral releases open weights for 1 trillion-parameter Large 4
The model adds reasoning controls after customer testing, while smaller versions, including one designed to run on a laptop, are planned.
Paris-based Mistral has made the weights for Mistral Large 4 available on Hugging Face, three weeks after previewing the 1 trillion-parameter AI model. The company says additional reinforcement learning informed by customer feedback improved its coding and cyber defense capabilities.
Large 4 accepts images and text as inputs and generates code and text. It activates 52 billion parameters per task. Mistral is also releasing a quantized version that can run on four Nvidia B200 or B300 GPUs.
The model's Terminal-Bench 4.0 score rose to 41% from 28% in the preview, compared with 40% for GLM-5.3. On Artificial Analysis’ Coding Agent Index, it rivals GLM-5.3 while averaging 47% fewer tokens for each task.
On CyberGym-E2E, which tests the identification, reproduction and repair of memory-safety bugs, Large 4 scored 79%. That matched MiMo-V2.6-Pro, while DeepSeek V4.1 Flash scored 76%.
Mistral said government and enterprise testers requested greater control over reasoning time. It added low and medium settings alongside the high setting, which it said is now 50% more efficient. Some customers already use the model to test their own applications for security vulnerabilities.
Access also includes Mistral’s API and some Vibe users. Smaller distilled models are planned for the coming months, including a laptop-compatible version, but Mistral gave no firm dates. It has not published API pricing.
