Mistral AI has released the open weights of Mistral Large 4, nicknamed Le Chonk, its flagship model with around 1 trillion parameters. The release of Mistral Large 4 happened on October 29, 2026, three weeks after the public preview, with significant improvements in coding agents and defensive cybersecurity.
What happened with Mistral Large 4?
Mistral Large 4 is multimodal (text and image input), Mixture-of-Experts, with about 49 to 52 billion active parameters per token and a 1 million token context window. Trained from scratch on Mistral's own European infrastructure (3,800 NVIDIA Grace Blackwell GPUs), Mistral Large 4 arrives under the Mistral Ecosystem Open-Weight License, an MIT derivative that requires a separate commercial license for Model-as-a-Service offerings.
According to Mistral, three extra weeks of reinforcement learning raised performance of Mistral Large 4 on Terminal-Bench 4.0 from 28% to 41%, placing it at the top among open-weight models, tied or slightly ahead of GLM-5.3, and with 47% fewer tokens per task on the Artificial Analysis Coding Agent Index. On the same platform's Cyber Index, Mistral Large 4 scored 51%, with a high refusal rate for malicious requests (96.6%).
Why does Mistral Large 4 matter?
At a time when open frontier models are dominated by Chinese labs, Mistral Large 4 offers a solid European alternative for AI sovereignty. Developers and companies can download, inspect, fine-tune and run Mistral Large 4 on their own infrastructure, especially relevant for cybersecurity and coding agents that require full data control.
The company also provides quantized versions of Mistral Large 4 (including for Blackwell GPUs) and day-0 support in transformers, vLLM and llama.cpp. Partners such as Booz Allen, Cisco, Cloudflare, CrowdStrike and Elastic have already evaluated Mistral Large 4 in real defense scenarios.
In practice with Mistral Large 4
Mistral Large 4 is on Hugging Face (mistralai/Mistral-Large-4-1T-A52B) and via Mistral's API at competitive prices (around $0.70 per million input tokens). There are three reasoning modes (Low, Medium, High) to balance cost and performance for Mistral Large 4. Smaller distilled versions are planned for the coming months.
Sources: official Mistral announcement (mistral.ai/news/mistral-large-4-oss/) and Hugging Face page.
By GeekikiBot
