Mistral AI has made its largest model downloadable. On October 29 the company released the weights of Mistral Large 4, a natively multimodal model it calls Le Chonk, after a public preview that began earlier in the month. The version now on Hugging Face is not the preview checkpoint: Mistral says it spent about three more weeks on reinforcement learning, mostly aimed at software engineering and defensive cybersecurity, and that the new build also reasons more efficiently.
According to Hugging Face's technical write-up, the model is a sparse Mixture-of-Experts design with roughly 1 trillion total parameters, about 49 billion of them active for each token, and a context window of 1 million tokens. A router picks four of 128 experts per token, alongside one shared expert, and a 1.6 billion-parameter Pixtral encoder handles images. Input can be text or pictures; output is text. Developers can now choose low, medium or high reasoning effort, with high as the default.
Mistral's headline claims are about code. It says Mistral Large 4 scores 48% on the Artificial Analysis Coding Agent Index, level with GLM-5.3 as the strongest open-weight entry, while spending 47% fewer tokens per task. The company also reports 41% on Terminal-Bench 4.0, a 13-point gain over the preview, and a 19.7% lead on Cognition's private FrontierCode 2.0 test. Hugging Face's own post quotes a slightly different Coding Agent Index figure of 49.8%, a reminder that these numbers come from the vendor and its launch partners and have not been reproduced by outside researchers.
Security is the second pillar. Mistral reports a 51% score on Artificial Analysis's Cyber Index, a test of whether a model can find, reproduce and patch vulnerabilities, and says the model refused 96.6% of malicious cyber prompts across three jailbreak benchmarks. It quotes testers including Booz Allen, Cisco, CrowdStrike, Cloudflare and the EU cybersecurity agency ENISA. The company is also candid about risk: after watching live traffic during the preview, it says it saw potential uplift for real attacks and is tightening its guardrails. A more permissive variant built for penetration testing is limited to vetted partners such as Orange Cyberdefense.
The license matters as much as the benchmarks. Mistral Large 4 comes with the Mistral Ecosystem Open-Weight license, which adapts the MIT text and allows use, modification, fine-tuning and commercial deployment. The single carve-out targets businesses that resell the model as a hosted service, which need a separate commercial agreement. That keeps the weights open for enterprises and researchers while protecting Mistral's own API business.
Distribution is broad from day one. The model runs through Mistral Studio and the Mistral API at $0.70 per million input tokens and $2.10 per million output tokens, on OpenRouter, in Microsoft Foundry and as an NVIDIA NIM container. Hugging Face lists day-zero support in transformers, vLLM and llama.cpp, plus quantized GGUF files; its engineers report that a 2-bit version fits on a single Apple Silicon machine with 512 GB of memory, while other formats need several GPUs.
Mistral frames the release around sovereignty, stressing that the model was trained from scratch on European infrastructure. Whether Le Chonk holds its place once independent evaluators test the final weights, and how the tightened cyber guardrails behave in self-hosted deployments that Mistral no longer controls, are the open questions.