Mistral AI announced the release of the weights for its Mistral Large 4 (ML4) model, known as "Le Chonk," on October 29, 2026. This trillion-parameter, multimodal mixture-of-experts model is now available for download on Hugging Face and through Mistral's API. The company states that ML4 offers advanced capabilities for agentic coding and defensive cybersecurity, positioning it as a leading open-weight solution.
The release follows a public preview of ML4 on October 6, 2026. Since then, Mistral has incorporated customer feedback and scaled reinforcement learning to improve the model's performance in coding and cybersecurity. Pierre Stock, Vice President of Science at Mistral, stated that the open-weight version of ML4 demonstrates the science team's ability to innovate rapidly, with continuous gains and no signs of saturation.
Mistral claims ML4 is the best open agentic coding model globally. The model achieved a 48% score on the Artificial Analysis Coding Agent Index, matching leading open-weight models like GLM-5.3 while using 47% fewer tokens per task on average. On the Terminal-Bench 4.0, which evaluates AI agents' coding abilities in a terminal, ML4's score increased from 28% to 41% after additional training, leading among open-weight models.
For cybersecurity, Mistral asserts that ML4 provides "frontier AI cyber defense capabilities" that enterprises and governments can fully own and control. This includes the ability to detect, analyze, and respond to cyber threats across code and IT infrastructure. ML4 scored 51% on the Artificial Analysis Cyber Index, which measures a model's ability to identify, reproduce, and patch code vulnerabilities without disrupting existing functionality. The model also reportedly refuses 96.6% of malicious cyber requests and resists 97.1% of attacks on the B3 Agent Security Benchmark. However, some independent analysis indicates that ML4 ranks third among open models in cyber defense, behind DeepSeek V4.1 Flash and MiMo-V2.6-Pro.
ML4 operates on a mixture-of-experts architecture with 1.05 trillion parameters in total, activating approximately 52 billion parameters per task for efficient processing. It also offers three reasoning modes, low, medium, and high, allowing users to adjust the depth of reasoning based on the complexity of the task. The model is natively multimodal, capable of processing both text and images. Mistral also highlights ML4's ability to convert 2D blueprints into 3D models using computer-aided design software, analyze financial data, and use satellite imagery for wildfire detection and crop monitoring.
The weights for Mistral Large 4 are released under a modified MIT license, the Mistral Ecosystem Open-Weight License. This license permits free use, modification, fine-tuning, and commercialization, with a specific condition: businesses offering a Model-as-a-Service must obtain a separate commercial license from Mistral. API access for ML4 is priced at $0.70 per million input tokens and $2.10 per million output tokens. Mistral states that ML4 was developed and trained entirely on European infrastructure.