Mistral Releases Open Weights for Large 4 'Le Chonk', a 1-Trillion-Parameter Model
Mistral says the mixture-of-experts model, now downloadable under a modified MIT license, is the best open-weight model for coding; its API costs $0.70 per million input tokens.
Published
Last updated / Fact-checked on
Mistral AI released the open weights of Mistral Large 4, nicknamed "Le Chonk," on October 29, 2026. The model is a roughly 1-trillion-parameter mixture-of-experts system with a 1-million-token context window. Mistral calls it the best open-weight model for coding, and its API is priced at $0.70 per million input tokens and $2.10 per million output tokens.
Illustrative photo of the Paris skyline. Image: Anthony Delanoix (Unsplash) / CC0 1.0, via Wikimedia Commons
From preview to open weights
Mistral first offered Large 4 as a public preview earlier in October; Euronews dated that launch to October 6. The company says three more weeks of reinforcement learning since then brought large gains in coding and defensive cybersecurity, plus better reasoning efficiency that lowers inference costs.
Hugging Face's technical write-up says the model has about 1 trillion total parameters, with about 49 billion active per token, and a 1-million-token context window. It accepts text and images and produces text, using a Pixtral vision encoder. Mistral says the model was trained from scratch on European infrastructure.
Coding results, by Mistral's numbers
All of the following benchmark results are reported by Mistral:
| Benchmark | Mistral Large 4 result |
|---|---|
| Cognition FrontierCode 2.0 (private) | 19.7%, top score |
| Artificial Analysis Coding Agent Index | 48%, effectively tied with GLM-5.3 while using 47% fewer tokens |
| Terminal-Bench 4.0 | 41%, up 13 points from preview, with 30% fewer output tokens |
| Artificial Analysis Cyber Index | 51% |
| B3 Agent Security Benchmark | Resisted 97.1% of attacks, up from 93.3% in preview |
Mistral also cites Surge Code's blind pairwise tests, in which graders preferred Large 4 over GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash. Only Claude Opus 5.5 ranked higher among the models tested.
These numbers come from Mistral and from benchmarks it selected. Rankings depend on the test, and independent results may differ. Hugging Face's own write-up, for example, reports a Coding Agent Index score of 49.8%.
The model offers three reasoning modes: low for simple tasks, medium for everyday work and high, the default, for the hardest problems.
A pitch to defenders
Mistral is aiming the release squarely at cybersecurity teams, a focus Reuters also highlighted. The company says the model can find, reproduce and patch vulnerabilities without refusing legitimate defensive work. Its guardrails also refused malicious cyber requests 96.6% of the time on average, which Mistral says is the highest rate it measured among open models.
Mistral also admits limits. While monitoring traffic during the preview, it saw the model refuse black-box attacks on live targets, but it said there is "a potential for uplift to conduct real attacks." It is tightening its guardrails as a result. A specialized variant for penetration testing and red-teaming is available only to vetted partners such as Orange Cyberdefense.
License, access and price
The weights are on Hugging Face and run on vLLM or SGLang. The model is also available through Mistral Studio, the Mistral API, OpenRouter, NVIDIA endpoints and Microsoft Foundry.
It uses the Mistral Ecosystem Open-Weight License, a modified MIT license. Anyone can use, modify, fine-tune and commercialize the model. The one exception is companies that run a model-as-a-service offering, which need a separate commercial license.
API pricing per million tokens:
- Input: $0.70
- Cached input: $0.07
- Output: $2.10
Why it matters
Large 4 gives organizations a frontier-scale model they can host and audit themselves. That matters most for banks, governments and security teams that cannot send data to an outside API. The trade-off is hardware: Hugging Face's recipes for the full model run on eight-GPU servers, so most smaller teams will use the API.
If you are evaluating it, run it against your own repositories and security tasks rather than relying on vendor benchmarks. Check the model-as-a-service clause before building a hosted product on top of it.
More model coverage is in our Artificial Intelligence hub.
How we reported this
We checked the specifications, benchmarks, license and pricing against Mistral's launch post. We confirmed the architecture through Hugging Face's technical blog, the preview date through Euronews and the cybersecurity focus through Reuters. Benchmark results are labeled as Mistral's figures. We removed the draft's claim that prices were cut, because Mistral only stated current prices.
Sources
- Mistral AI: Mistral Large 4 (Le Chonk) is now open weight, Oct 29, 2026.
- Hugging Face: Welcome Le Chonk, Mistral Large 4 on Hugging Face, Oct 29, 2026.
- Reuters via Investing.com: Mistral AI publicly releases Large 4, targets cybersecurity users, Oct 29, 2026.
- Euronews: What to know about Mistral's ML4, Oct 6, 2026.
DailyMech used AI tools to help research and draft this article; it was fact-checked by our editors.