Mistral Large 4 open weights release date and technical details

On October 29, French artificial intelligence company Mistral AI announced the release and public preview exit of its new flagship model, Mistral Large 4. Internally nicknamed "Le Chonk" due to its massive one-trillion-parameter total size, the model is designed for self-hosting, advanced coding, agentic workflows, and robust information security.

• The model totals one trillion parameters, but uses a Mixture of Experts (MoE) architecture where only about 49 to 52 billion parameters are active per token.
• Model weights are publicly available on Hugging Face without requiring manual access requests.
• It offers high efficiency, consuming up to 47% fewer tokens per task than competitors while maintaining competitive accuracy.

Architecture and Technical Specs Built on a Mixture of Experts architecture featuring 128 routed experts with four active per token plus one shared expert, the model underwent initial training for roughly two months on approximately 3,800 NVIDIA Grace Blackwell GPUs in European datacenters. Following this, Mistral spent three weeks applying reinforcement learning on its own infrastructure in France, significantly enhancing the model's performance on agent benchmarks. The context window supports one million tokens, and the model accepts both text and image inputs via a Pixtral-style encoder. Additionally, new low and medium reasoning modes have been added to give clients flexible control over computational cost and execution time for simpler tasks.

Performance and Hosting Requirements Despite being an open-weight release, running "Le Chonk" locally requires substantial enterprise hardware rather than a standard consumer PC. Deploying the NVFP4 quantized version for Blackwell architecture GPUs requires around 610 GB of disk space, while standard versions demand even greater resources. Although OpenAI CEO Sam Altman has criticized local models for inefficient hardware utilization, Mistral remains committed to fostering independent European AI infrastructure. On benchmarks, Mistral Large 4 shows impressive improvements, jumping from 28% to 41% on Terminal-Bench 4.0 and reaching 84.5% on BrowseComp. Security is a major focal point: the model successfully rejects 96.6% of malicious prompts and withstands 97% of attacks in the B3 Agent Security benchmark, positioning it as a highly secure enterprise option.

Industry Impact and Licensing The launch of Mistral Large 4 reinforces Europe's standing in the foundational AI landscape alongside dominant US and Chinese labs. Distributed under the modified Mistral Ecosystem Open-Weight License, the model permits commercial use and fine-tuning, though providing access as a managed API service requires a commercial agreement with Mistral. This gives businesses a powerful tool to maintain data sovereignty while building tailored AI solutions.