Mistral AI released Mistral Large 4, nicknamed “Le Chonk,” on October 29, describing it as an open-weight, multimodal and multilingual model with 1 trillion total parameters, 49 billion active parameters and a context window of up to 1 million tokens. It accepts text and images and generates text; downloadable weights are available on Hugging Face, alongside API access and partner platforms.

Mistral says additional reinforcement learning since the public preview strengthened coding, reasoning efficiency, defensive cybersecurity and other enterprise capabilities, while noting that training is ongoing and the preview may not reflect the model’s final capabilities.

The company presents open weights as a route to sovereign AI: organizations can inspect, adapt and run the model on their own infrastructure, an option it says may appeal to sensitive sectors. Mistral says it trained the model from scratch in Europe on 3,800 NVIDIA Grace Blackwell GPUs and plans independently operated deployments in Europe under European law; managed and private options are also available through Microsoft Foundry.

Under the hood, Mistral Large 4 uses a sparse mixture-of-experts architecture, with 71 MoE layers routing each token to four of 128 experts plus a shared expert. Its 72-layer text decoder combines Multi-head Latent Attention with a sparse attention indexer, while the model also includes a vision encoder for image processing.

Developers can choose Low, Medium or High reasoning modes to balance performance and cost, with High set as the default. The model is positioned for complex software engineering and agent workflows, as well as finance, legal work, scientific coding and visual analysis of charts, technical drawings and long documents.

Mistral highlights coding and agentic-work results, including a 19.7% score on Cognition’s FrontierCode 2.0 benchmark and a 49.8% score on the Artificial Analysis Coding Agent Index. It also reports 41% on Terminal-Bench 4.0, a 13-point improvement over its preview result, and says the model used 47% fewer tokens per task than GLM-5.3 in a comparison.

Cybersecurity is a central part of the pitch: Mistral reports an 82% score on a vulnerability reproduction-and-patching test, 93% of Cybench challenges solved and a 93.3% score on Lakera’s B3 benchmark. It also cites a 51% score on the Artificial Analysis Cyber Index and a 96.6% average refusal rate on malicious cyber requests, while acknowledging that live-traffic monitoring surfaced potential misuse risks and that guardrails are being revised.

Vetted enterprise partners, including Orange Cyberdefense, are previewing a specialized penetration-testing variant under contractual safeguards that limit testing to owned or explicitly authorized systems. Customers can also tailor the model with proprietary security data and policies using Mistral Forge.

Mistral says organizations can use the model to detect, analyze and respond to threats across code and IT infrastructure, including through on-premises deployment and coordinated responses to major incidents. Co-founder Guillaume Lample said open models could help organizations defend against attackers using jailbroken closed models.

Mistral Large 4 is available through Mistral Studio and API, OpenRouter, Microsoft Foundry and NVIDIA services, as well as for self-hosting; it also has day-one support in Transformers, vLLM and llama.cpp, and can be fine-tuned using TRL. The modified MIT-style Mistral Ecosystem Open-Weight License permits use, modification, fine-tuning and commercialization, but businesses offering the model as a service need a separate commercial license.

Mistral lists API pricing at $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens. The weights are offered in FP8 and NVFP4 versions, and Mistral measured a speculative-decoding head included in the checkpoint at up to 1.9 times the throughput of serving the model alone.

The release advances Mistral’s argument that organizations, particularly those handling cybersecurity, should control their AI systems and data rather than depend on providers such as OpenAI or Anthropic. CEO Arthur Mensch described AI as “a tool for empowerment, not dominance,” and Mistral says Large 4 will serve as a foundation for future specialized and optimized models.