French AI company Mistral has released the open weights of Mistral Large 4, the flagship model nicknamed Le Chonk, giving developers, companies and governments the chance to download and run one of the largest openly available models themselves. The release follows a public preview in early October, after which Mistral says three more weeks of reinforcement learning lifted the model well past its preview performance.

The model is a natively multimodal mixture-of-experts system with about 1 trillion total parameters, of which roughly 49 billion are active for each token, or 52 billion when embeddings are counted. It accepts text and images as input and produces text, and it offers a context window of 1 million tokens. Mistral has published the weights on Hugging Face as mistralai/Mistral-Large-4-1T-A52B, in FP8 and NVFP4 variants, alongside access through its own API and Mistral Studio.

Built for coding agents

Mistral is positioning Large 4 first and foremost as a model for agentic coding. On Cognition's FrontierCode 2.0 benchmark, which tests whether a repository maintainer would accept an AI agent's code change, the company reports a score of 19.7 percent, putting it ahead of rival open-weight models. On Terminal-Bench 4.0, which measures terminal-based coding work, it reports 41 percent, up 13 percentage points from the preview version.

On the Artificial Analysis Coding Agent Index, Mistral Large 4 scores 48 percent, effectively tied with the leading open-weight model, GLM-5.3, while using 47 percent fewer tokens per task on average, according to Mistral. In blind human-preference tests run by Surge Code, the model was preferred over leading open-weight rivals including GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash, ranking behind only Claude Opus 5.5 among the models evaluated. These are largely vendor-reported figures, and independent evaluations are still catching up, but the direction is consistent: Mistral is competing at the top of the open-weight field.

Cybersecurity focus

The second pillar of the release is defensive cybersecurity. Mistral says the model scores 51 percent on the Artificial Analysis Cyber Index, which measures the ability to find vulnerabilities in code, reproduce and validate them, and patch them without breaking existing functions. The company also reports that it refuses an average of 96.6 percent of malicious cyber requests across JailbreakBench, StrongREJECT and AgentHarm, which it describes as the highest refusal rate it has measured among open models.

Mistral says it worked with leading cybersecurity vendors to test the model in real-world settings, and it names Booz Allen Hamilton, Cisco, Cloudflare, CrowdStrike, Elastic, Fortinet and the EU cybersecurity agency ENISA among the organisations evaluating or endorsing it. Enterprise design partners such as Orange Cyberdefense are getting preview access to a specialised variant tuned for penetration testing, vulnerability assessment and red-teaming, subject to vetting and contractual safeguards.

Three reasoning modes and sovereign AI

To give developers more control over the trade-off between performance and cost, Large 4 ships with three reasoning modes: low for simple tasks, medium for everyday work, and high, the default, for the hardest cases. Mistral also reports gains beyond coding and security, including 84.5 percent on the BrowseComp web-browsing benchmark and 73.3 percent on Finch, a finance workflow benchmark.

The model was trained from scratch on European infrastructure, and Mistral is framing the release as a step towards sovereign AI, giving organisations a model they can host, customise and audit under their own control. The weights are published under the Mistral Ecosystem Open-Weight License, a modified MIT licence that allows free use and fine-tuning, although running a model-as-a-service business requires a separate commercial licence from Mistral.

Self-hosting is not a light undertaking. The full FP8 checkpoint is about 1.06 terabytes, while the smaller NVFP4 version is around 610 gigabytes, and Mistral's deployment guidance targets servers with eight GPUs. For most users, the API is the practical route, priced at 0.70 dollars per million input tokens, 0.07 dollars for cached input and 2.10 dollars per million output tokens. A full technical report is still to come, according to Mistral, so the benchmark claims should be read as the company's own account until independent testing lands.