Mistral Releases Large 4 Open Weights: 1-Trillion-Parameter Flagship Leads Open Models in Coding and Cyber Defense

What happened

On Oct 29, Mistral announced in its official blog that it had released the weights of Large 4 (Le Chonk): the model entered public preview on Oct 6, and three weeks of additional reinforcement learning on feedback from government and corporate testers improved its coding and defensive cybersecurity capabilities while raising inference efficiency.

Reuters reported the same day that Mistral described it as its most capable openly available model for cybersecurity use, downloadable and runnable independently so companies can use it on their own systems instead of depending on a third-party provider; Mistral Vice President of Science Pierre Stock said he was not comfortable living in a world without open weights, and that once weights are out, access cannot be contained.

Mistral is also giving enterprise design partners, including Orange Cyberdefense, preview access to a specialized Large 4 variant for cybersecurity workflows such as penetration testing, vulnerability assessment and red-teaming, subject to rigorous vetting and contractual safeguards, with testing restricted to owned or explicitly authorized systems.

Key facts

Scale: Large 4 is a 1-trillion-parameter, natively multimodal flagship model, trained from scratch in Europe.

Access: weights are downloadable from Hugging Face for self-deployment with vLLM or SGLang; the model is also available via Mistral Studio, the Mistral API and OpenRouter, with NVIDIA NIM containers and Microsoft Foundry offering it too.

API pricing: from today, $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens.

License: the Mistral Ecosystem Open-Weight license, a modified MIT license, allows free use, modification, fine-tuning and commercialization; businesses operating a model-as-a-service offering must obtain a separate commercial license from Mistral.

Company-reported benchmarks: 19.7% on FrontierCode 2.0, leading open-weight models; 48% on the Artificial Analysis Coding Agent Index, tied with GLM-5.3 while using 47% fewer tokens per task on average; 41% on Terminal-Bench 4.0, up 13 percentage points from the preview; 51% on the Cyber Index, with a 96.6% average refusal rate on malicious cyber requests.

Reasoning modes: three new modes, Low, Medium and High, for simple tasks, everyday tasks and the hardest use cases respectively, with High as the default.

Timeline: public preview on Oct 6 with weights promised by the end of October; weights officially released on Oct 29.