Mistral Large 4 Open Weights Released: 1 Trillion Parameters
Mistral AI has published open weights for Mistral Large 4, a trillion-parameter Mixture of Experts model nicknamed «Le Chonk», after a three-week reinforcement learning push.
Mistral AI released the open weights of Mistral Large 4 on October 29. Inside the company the model is nicknamed «Le Chonk» because of its size: one trillion parameters. The model has now left public preview, with improvements in coding, agentic tasks and security.
The weights are available on Hugging Face under mistralai/Mistral-Large-4-1T-A52B, with no access request required. A quantized NVFP4 build for NVIDIA Blackwell GPUs is also available. The model was first shown on October 6 as an API-only preview, and Mistral kept its promise to publish weights within a month. In the meantime it applied reinforcement learning and tested the model with customers. According to science VP Pierre Stock, that stage ran on the company's own French infrastructure drawing about 10 MW, while base training took roughly two months on about 3,800 NVIDIA Grace Blackwell GPUs.
Architecture and benchmarks
The model uses 128 routed experts, four of them active per token, plus one shared expert. Around 49 billion parameters are active, or 52 billion including embeddings and the output layer, which explains the repository name. The context window is 1 million tokens, and the model accepts text and images.
The three weeks of tuning show most clearly in agentic tests. Terminal-Bench 4.0 rose from 28% to 41% with 30% fewer output tokens, and BrowseComp reached 84.5% versus 75% for the SFT checkpoint. Low and medium reasoning modes were added alongside high. Mistral also says the model rejects 96.6% of harmful requests and withstands 97% of attacks in the B3 Agent Security benchmark. It uses 47% fewer tokens per task than GLM-5.3 at slightly lower accuracy, and its coding lead over GLM is about one point. All figures come from Mistral; only results via Artificial Analysis indexes are independently sourced, and even those appear in the company's announcement.
Hardware and licence limits
Open weights do not mean easy deployment. The FP8 version needs about 1 TB of storage, and NVFP4 about 610 GB, and it requires Blackwell-generation GPUs. The licence, a modified MIT variant called the Mistral Ecosystem Open-Weight License, allows use, fine-tuning and commercial application, but offering the model as a service requires a commercial licence from Mistral.
The release comes as OpenAI chief Sam Altman called local models a «terrible allocation of scarce resources», arguing hardware belongs in data centres. Mistral, the EU's main foundation-model developer, is betting on the opposite approach.
