Mistral Large 4 (Le Chonk) release notes

Overview of Mistral Large 4 (Le Chonk)

Mistral Large 4, codenamed "Le Chonk," is a 1-trillion-parameter, natively multimodal model designed for agentic coding and defensive cybersecurity. Following a public preview, the model has undergone three weeks of additional reinforcement learning, resulting in significant performance gains in coding benchmarks and reasoning efficiency. It is available as an open-weight model, supporting sovereign AI strategies by allowing organizations to self-host and customize the technology on their own infrastructure.

Coding and Agentic Performance

Mistral Large 4 has established itself as a leading open-weight model for coding tasks. On the FrontierCode 2.0 benchmark, the model achieved a score of 19.7%. It demonstrates high token efficiency, scoring 48% on the Artificial Analysis Coding Agent Index while using 47% fewer tokens per task than comparable models. Additionally, the model improved its performance on Terminal-Bench 4.0 by 13 percentage points, reaching 41% while reducing output tokens by 30%. To manage the balance between cost and performance, the model supports three distinct reasoning modes: Low, Medium, and High.

Cybersecurity Capabilities

Designed for defensive security, Mistral Large 4 is capable of vulnerability reproduction, validation, and patching. It scored 51% on the Artificial Analysis Cyber Index. The model includes built-in guardrails that distinguish between defensive tasks and malicious requests, achieving a 96.6% refusal rate on harmful cyber prompts. For enterprise partners, Mistral provides a specialized variant of the model optimized for advanced security workflows, including penetration testing and red-teaming, subject to rigorous vetting and contractual safeguards.

Enterprise and Operational Metrics

Beyond coding and security, Mistral Large 4 shows improvements in general enterprise workflows. It achieved a 73.3% score on the FinWorkBench (Finch) for finance-related tasks and 84.5% on the BrowseComp benchmark for web browsing. The model also features enhanced security for agentic deployments, resisting 97.1% of attacks on the B3 Agent Security Benchmark, which protects against threats like system prompt exfiltration and malicious code injection.

Deployment and Sovereignty

Organizations can deploy Mistral Large 4 on-premises or via private cloud environments to maintain control over their data and intelligence. The model is available via the Mistral API, Mistral Studio, and OpenRouter, with open weights downloadable from Hugging Face. For enterprise-scale operations, it is accessible through Microsoft Foundry and NVIDIA-hosted endpoints. The model is released under the Mistral Ecosystem Open-Weight license, a modified MIT license that permits free use, modification, and commercialization, with a separate commercial license required for entities operating a Model-as-a-Service (MaaS) offering.