Mistral Releases Open Weights for Its Trillion-Parameter Large 4 Model

Mistral AI said on 29 October that it is releasing the weights of Mistral Large 4. The company calls the model Le Chonk. It is a hybrid instruct-and-reasoning mixture-of-experts model with multimodal input, trained from scratch on European infrastructure. Mistral says this version improves on the public preview it released earlier in October.[1][2][3]

Open weights mean the trained parameters can be downloaded and run on an organisation’s own servers. Mistral names Hugging Face for the download and vLLM or SGLang for deployment. It also sells hosted access through Mistral Studio, its API and OpenRouter, at US$0.70 per million input tokens and US$2.10 per million output tokens.[1]

Mistral says the model is free to use, modify, fine-tune and commercialise under the Mistral Ecosystem Open-Weight licence, which it describes as a modified MIT licence. There is one condition. A business that operates a Model-as-a-Service offering must get a separate commercial licence from Mistral before using the model commercially. We read Mistral’s description of the licence, not the licence text itself.[1]

That condition matters more to cloud and platform firms than to ordinary users. A Malaysian company that fine-tunes the model for its own staff or customers is not obviously running a Model-as-a-Service business. A firm that resells model access probably is. Mistral’s post does not define the term, so a lawyer’s opinion is sensible before launching a paid product.[1]

Mistral says Large 4 is the best open-weight model for coding. It reports 19.7% on Cognition’s private FrontierCode 2.0 benchmark and 41% on Terminal-Bench 4.0, up 13 points from the preview. On the Artificial Analysis Coding Agent Index it scores 48%, roughly tied with GLM-5.3 while using 47% fewer tokens per task. In Surge Code’s blind tests it ranked behind only Claude Opus 5.5.[1]

These are vendor-reported results, and some rest on private or third-party benchmarks. Mistral also reports 51% on the Artificial Analysis Cyber Index, a 96.6% average refusal rate on malicious cyber requests and 97.1% attack resistance on the B3 agent-security benchmark. It adds that live traffic after the preview showed potential for uplift in real attacks, so it is changing its guardrails.[1]

Mistral lists several routes. Teams can use Mistral Studio, the Mistral API or OpenRouter, or download the weights and self-host them. NVIDIA offers hosted endpoints and a NIM container, and Microsoft says Mistral Large 4 is available through Microsoft Foundry, with private deployment options for Azure developers. Mistral also offers customisation through its Forge service.[1]

The preview post gives the scale. Large 4 has 1 trillion parameters, of which 52 billion are active, and Mistral trained it on 3,800 NVIDIA Grace Blackwell GPUs in its own European data centres. The weights post gives no memory or GPU requirement for self-hosting.[2][1]