Mistral AI has released its latest large language model, Mistral Large 4, which is currently available in public preview via the company's cloud platform. The model utilizes a mixture-of-experts architecture with 1 trillion parameters, activating 49 billion parameters per prompt. It supports over 160 languages and demonstrated an 82% score on the AA Cyber Index for vulnerability patching.
In computer vision tasks, the model outperformed GPT-6 Astra by 1% on the Dense200 benchmark. While it trails behind frontier models in coding benchmarks, it surpassed open-source competitors such as Qwen3.8 Max and DeepSeek V4 Pro on the AutomationBench and AA-Briefcase tests. The company plans to release the model weights later this month.
The training process for Mistral Large 4 involved 3,800 Grace Blackwell chips, each consisting of two Nvidia Blackwell GPUs and one CPU. Engineers developed a specialized software stack to execute tens of thousands of parallel rollouts, generating 33 billion tokens daily. This asynchronous workflow allowed for continuous refinement and faster processing.
Mistral intends to continue this training workflow to develop more powerful versions of the model in the coming months. The company aims to leverage the architecture of Mistral Large 4 to create a series of specialized models tailored for specific use cases in the future.
