Mistral Large 4 (Le Chonk) open weights: what’s new.

Mistral released the open weights of Mistral Large 4, the model it nicknames Le Chonk, on 29 October 2026. It is an improved version of the model Mistral put into public preview earlier in October, and Mistral calls it the best open-weight model for coding. Here is what Mistral has confirmed about the Mistral Large 4 open weights release, what changed since the preview, what the licence allows, and how to try it.

What is Mistral Large 4 (Le Chonk)?

Mistral Large 4 is Mistral AI’s flagship model: a mixture-of-experts model that takes text and images as input and writes text. Mistral’s documentation lists 1.05 trillion total parameters, 52 billion active parameters, a 1.6 billion parameter vision encoder and a context window of 1 million tokens. Hugging Face’s launch post describes it as having about 1 trillion total and 49 billion active parameters.

Mistral describes it as a hybrid instruct-and-reasoning model that handles instructions, reasoning and agentic work in one model, and says it is natively fluent in more than 160 languages. Mistral also stresses that it was trained from scratch on European infrastructure, which it pitches to organisations that want what it calls sovereign AI: a model they can inspect, customise and run on their own hardware.

What’s new in the open-weight release

Mistral says it spent the weeks since the preview scaling reinforcement learning, which it credits for large gains in coding and defensive cybersecurity, plus better reasoning efficiency that lowers inference costs.

Coding

These are Mistral’s own reported results for the open-weight model:

• FrontierCode 2.0: 19.7% on Cognition’s private benchmark, which judges whether a repository maintainer would merge the agent’s change. Mistral says this leads the field it compared.
• Artificial Analysis Coding Agent Index: 48%, which Mistral calls effectively tied with the leading open-weight model, GLM-5.3, while using 47% fewer tokens per task on average.
• Terminal-Bench 4.0: 41%, up 13 percentage points from the preview, with 30% fewer output tokens.
• Human preference: in Surge Code’s blind pairwise evaluations, Mistral says graders preferred Large 4 over GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash, ranking it behind only Claude Opus 5.5 among the models evaluated.

The model offers three reasoning effort levels, low, medium and high, which Mistral says let developers balance cost and performance. According to Hugging Face, high is the default and reasoning cannot be switched off entirely.

Cybersecurity

Mistral’s tagline for the release is “Built for Coders, Trusted by Defenders.” It reports a score of 51% on the Artificial Analysis Cyber Index, which tests finding, reproducing and patching vulnerabilities, and says the model handles legitimate defensive tasks without unnecessary refusals. Hugging Face adds that the model scores 82% on the index’s test of reproducing a real vulnerability in open-source software and then patching it, and solves 93% of the 40 Cybench challenges.

On safety, Mistral reports an average refusal rate of 96.6% on malicious cyber requests from JailbreakBench, StrongREJECT and AgentHarm, the highest it measured among open models. Mistral also says that, after monitoring live traffic since the preview, it sees potential for uplift in real attacks and is tightening its guardrails. A specialised variant for penetration testing and red-teaming is in preview only for vetted enterprise design partners such as Orange Cyberdefense.

Knowledge work and agents

• Finance: 73.3% on FinWorkBench (Finch), which Mistral says is second only to Kimi K3 among open models.
• Web browsing: 84.5% on BrowseComp, according to Mistral.
• Agent security: Mistral says the model resists 97.1% of attacks on the B3 Agent Security Benchmark, up from 93.3% in the preview.

Mistral’s own demos include Orbital Trace, a live 3D map of roughly 19,000 satellites and debris pieces that Mistral says the model built and debugged with no human help, and a quantum chemistry solver with a 3D page where you pull a molecule apart and watch the electrons regroup.

Licence and pricing

The weights are released under the Mistral Ecosystem Open-Weight licence, which Mistral describes as a modified MIT licence. Mistral says the model is free to use, modify, fine-tune and commercialise, with one condition: businesses that run a Model-as-a-Service offering need a separate commercial licence from Mistral first.

Through Mistral’s API, Mistral says the model costs $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens.

How to try Mistral Large 4

1. In the browser or API. Mistral says you can start in Mistral Studio, through the Mistral API or on OpenRouter. It also says you can try the model in Vibe Code on your own repository; Mistral’s assistant apps are listed in our directory as Mistral Vibe.
2. Download the weights. The official FP8 and NVFP4 checkpoints are on Hugging Face, and Hugging Face says the model is supported in Transformers, vLLM and llama.cpp, with fine-tuning through TRL. Mistral also points self-hosters to vLLM and SGLang.
3. Use a managed cloud. Mistral lists NVIDIA-hosted endpoints and an NVIDIA NIM container, and Microsoft Foundry with private deployment options for Azure.
4. Customise it. Enterprises can tailor the model to their own data with Mistral Forge.

This is a very large model to self-host. SGLang’s guide runs the FP8 checkpoint on eight NVIDIA B300 GPUs and the NVFP4 checkpoint on four GB300 GPUs. Hugging Face says the 2-bit GGUF quantisations run on a single Apple Silicon Ultra machine with 512 GB of unified memory, while the other formats need several machines or GPUs.

What people are building with it

The Atlas only lists builds whose creators name the model they used, and we don’t have any Mistral Large 4 builds yet. To compare it with the models Mistral benchmarked it against, browse what people have made with Claude Opus 5.5 and Kimi K3. We’ll add Le Chonk builds as creators share them on X.

Mistral Large 4 FAQ

Mistral says a full technical report is coming soon, and we’ll update this post when it lands.