On October 29, 2026, Mistral released the full weights of Mistral Large 4, the 1-trillion-parameter natively multimodal model it had opened as an API preview earlier in the month. The weights are on Hugging Face and deploy with vLLM or SGLang. They ship under the Mistral Ecosystem Open-Weight License, a modified MIT license that allows use, fine-tuning and commercial use, with one condition: businesses running a Model-as-a-Service offering need a separate commercial license.
Mistral says the release followed about three more weeks of reinforcement learning focused on coding and cybersecurity. Terminal-Bench 4.0 rose to 41% from the preview’s 28%, with 30% fewer output tokens. Mistral also reports 19.7% on Cognition’s private FrontierCode 2.0, which it calls the leading open-weight score, 48% on the Artificial Analysis Coding Agent Index (effectively tied with GLM-5.3) and 84.5% on BrowseComp. Reasoning now has three modes: Low, Medium and High (the default).
The model is also hosted on Mistral Studio, the Mistral API, OpenRouter, NVIDIA endpoints and Microsoft Foundry, priced at $0.70 per million input tokens and $2.10 per million output. A specialized cybersecurity variant for penetration testing and red-teaming is in restricted preview for vetted design partners such as Orange Cyberdefense. Mistral says it updated cyber guardrails after watching live preview traffic and reports a 96.6% average refusal rate on malicious cyber requests.
Why it matters: a European lab has put a trillion-parameter model with competitive coding scores into open weights, while splitting off its most offensive cyber capability into a gated variant - an emerging template for releasing large open models.
What it does not show: the scores are Mistral’s or its partners’, and the announcement itself shows Kimi K3 ahead on BrowseComp and FinWorkBench. The license is not a plain open-source license because of the managed-service carve-out.