Mistral AI released the weights for its Mistral Large 4 AI model on October 29, 2026, giving businesses and governments a trillion-parameter system they can download, customize and run themselves. Nicknamed “Le Chonk,” the multimodal model leads open-weight coding benchmarks, according to Mistral, while combining defensive cybersecurity capabilities with lower token use.
Key takeaways
Le Chonk was trained from scratch on European infrastructure.
Mistral reports open-weight leads on FrontierCode and Terminal-Bench.
Commercial Model-as-a-Service providers need a separate license.
Mistral Large 4 brings a European-trained AI model to self-hosted deployments
Mistral Large 4 is a 1-trillion-parameter multimodal model whose released weights let organizations operate it without relying on Mistral’s API. It entered public preview on October 6, 2026, before the open-weight release on October 29.
Mistral says it trained the system from scratch on European infrastructure, with traceable origins from development through release. Organizations can inspect, modify and host the model themselves—capabilities the company places at the center of its sovereign AI offering.
Access is available through Mistral Studio, the Mistral API and OpenRouter. Developers can download the weights from Hugging Face and serve them using vLLM or SGLang. NVIDIA also offers hosted endpoints and an optimized NVIDIA NIM container.
Coding gains combine higher scores with fewer tokens
Mistral reports that Large 4 leads open-weight models on FrontierCode 2.0 and Terminal-Bench 4.0 , following three additional weeks of reinforcement learning after its preview. Cognition’s private FrontierCode benchmark, which assesses end-to-end code quality and the acceptability of proposed repository changes, gave it 19.7%.
On Terminal-Bench, it reached 41%, a gain of 13 percentage points over preview, while cutting output tokens by 30%. Its 48% score on the Artificial Analysis Coding Agent Index effectively matched GLM-5.3, using 47% fewer tokens per task on average.
Surge Code’s blind comparisons favored it over GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash. Among the evaluated models, only Claude Opus 5.5 ranked higher.
The cybersecurity results include 51% on the Artificial Analysis Cyber Index , covering vulnerability discovery, reproduction, validation and patching. Mistral also reports a 96.6% average refusal rate for malicious cybersecurity requests, the highest it measured among open models.
For enterprise work, the model scored 73.3% on Finch’s finance workflows, second to Kimi K3 among open models, and 84.5% on BrowseComp. Attack resistance on the B3 Agent Security Benchmark rose to 97.1%, from 93.3% in preview.
Customization, reasoning modes and the commercial license
The Mistral Large 4 AI model offers three reasoning modes: low, medium and high , allowing developers to adjust performance and cost. Low targets simple tasks, medium handles everyday work, and high—the default—targets demanding tasks requiring more thorough reasoning.
Enterprises can tailor the model to proprietary data, policies and security requirements using Mistral Forge, then build and govern applications through Mistral Studio. Deployment options include on-premise hosting and Microsoft Foundry on Azure, with private deployment options for Azure developers.
The Mistral Ecosystem Open-Weight license is a modified MIT license permitting free use, modification, fine-tuning and commercialization. Businesses offering Model-as-a-Service must obtain a separate commercial license before using it commercially.
API pricing per million tokens is $0.70 for input, $0.07 for cached input and $2.10 for output .
Security partners and La Banque Postale emphasize control
Mistral is giving enterprise design partners, including Orange Cyberdefense, preview access to a specialized Large 4 variant for penetration testing, vulnerability assessment and red-teaming. Access requires vetting and contractual safeguards, with tests limited to owned or explicitly authorized systems in controlled environments.
The company says it is updating its cybersecurity guardrails and working with customers to customize models around proprietary security data, detection logic and threat environments. Defined security tasks can also be optimized into smaller, specialized variants.
In an interview with TechCrunch, Mistral Vice President of Science Pierre Stock said open weights give organizations access to defensive AI that cannot be cut off by an outside provider.
La Banque Postale described the release as reinforcing its long-term partnership with Mistral, linking technological sovereignty to customer trust. Its stated ambition is to combine human expertise and digital technology to improve experiences for customers and employees.
Article produced with the assistance of artificial intelligence and reviewed by the editorial team.
