Mistral releases one-trillion-parameter model Le Chonk in open-weight preview

Model architecture and training

Paris-based artificial intelligence startup Mistral released an API preview of its one-trillion-parameter model, Mistral Large 4, on 6 October 2026. Nicknamed "Le Chonk" after an internet meme referencing large cats, the multimodal system activates 49 billion parameters during inference tasks. Mistral trained the system from scratch over a two-month period using approximately 4,000 Nvidia Grace Blackwell graphics processing units in its European data centres, consuming roughly 10 megawatts of power. The training dataset encompasses more than 160 languages, including every official language of the European Union. Reinforcement learning remains ongoing ahead of the public distribution of weights.

There are a lot of areas where the other labs will not focus that much. There are so many domains in which you can improve models.

Benchmark performance and vision capabilities

Mistral published preliminary evaluation results across coding, finance, legal analysis, and computer vision benchmarks. On DeepSWE, an agentic software engineering test, Mistral Large 4 recorded 62%, leading Zhipu's GLM-5.3 at 61%, DeepSeek-V4-Pro at 57%, Qwen 3.8 Max at 51%, and Reflection AI Beam at 44%. On FinWorkBench, the model achieved 67%, matching DeepSeek-V4-Pro and surpassing GLM-5.3 at 65% and Mistral Medium 3.5 at 37%. In the Harvey Legal Agent benchmark, it scored 15%, compared to 13% for Kimi K3 and 5% for OpenAI's GPT-6 Astra. On the DIOR-RSVG remote-sensing vision test, Mistral Large 4 recorded 73% against 68% for GPT-6 Astra, with both models scoring 42% on Dense200. Mistral plans to apply these visual capabilities to satellite damage analysis and forest fire detection.

Enterprise cyberdefense and sovereignty

Mistral positioned the architecture specifically toward enterprise clients and sovereign entities requiring self-hosted infrastructure, custom tuning, and zero-data-retention environments. The release emphasizes defensive cybersecurity applications following industry security incidents, with Mistral participating in Nvidia's Open Secure AI Alliance established after the Hugging Face breach. Mistral is providing developers, cybersecurity firms, and government agencies with an unconstrained preview version during a three-week evaluation window before releasing the model weights under a custom license on 27 October 2026.

The cyber defense capabilities will enable enterprises and governments to defend themselves against threat actors that are jailbreaking closed models to perform cyberattacks.

Commercial expansion and rollout schedule

The release follows rapid institutional growth and capital accumulation for the French AI company. Mistral expanded its research science division from three researchers to roughly 300 employees, while company earnings increased 20-fold over the preceding year through pay-as-you-go cloud hosting and dedicated engineering support. In September 2026, Mistral completed a $3.3 billion financing round at a $24 billion valuation. The name Le Chonk originated in June 2026, when a satirical social media meme about a fictional 30-trillion-parameter model called "Le Chaton Fat" circulated across Reddit and X before Mistral formally adopted the branding.