Mistral Drops Weights for 1T Parameter Mistral Large 4

On October 29, French AI lab Mistral AI made the weights for its flagship model, Mistral Large 4, publicly downloadable on Hugging Face. Internally nicknamed "Le Chonk", the model contains a massive one trillion parameters, though its Mixture of Experts (MoE) design keeps computational demand during inference to roughly 52 billion active parameters per token.

The release marks the end of a preview phase started on October 6. According to Mistral, the architecture routes each token through four of its 128 experts alongside one shared expert, supporting a 1-million-token context window with multimodal text and image inputs. Pre-training reportedly required two months across roughly 3,800 NVIDIA Grace Blackwell GPUs, followed by three weeks of reinforcement learning consuming around 10 MW at the company's French facilities.

Mistral claims the post-preview tuning boosted performance on Terminal-Bench 4.0 from 28% to 41% while cutting output token volume by 30%. The company highlights efficiency against Chinese rival GLM-5.3, reporting comparable coding output while consuming 47% fewer tokens per task, alongside high rejection rates on security benchmarks. However, all cited benchmark numbers originate directly from Mistral's own technical report rather than independent audits.

While available without access gating, running "Le Chonk" locally remains out of reach for consumer setups. Storing the FP8 build demands around 1 TB of storage, while the quantized NVFP4 release still requires roughly 610 GB and specialized NVIDIA Blackwell silicon. Commercial hosting faces another hurdle: the files use the Mistral Ecosystem Open-Weight License, which permits internal business deployment and fine-tuning but demands a separate commercial agreement for anyone offering the model as a service.

Source: habr