The short answer
● What shipped — Mistral Large 4, open weights
● Size — 1 trillion parameters, roughly 50B active
● Licence — Mistral Ecosystem Open-Weight, a modified MIT
● The catch — Model-as-a-Service needs a separate commercial licence
● API price — $0.70 in, $0.07 cached, $2.10 out per million
Mistral released the weights for Mistral Large 4, which it calls Le Chonk: a natively multimodal mixture of experts at a trillion total parameters, trained from scratch on European infrastructure. The weights are on Hugging Face, with NVIDIA NIM containers, Microsoft Foundry and OpenRouter alongside Mistral’s own API.
On benchmarks it posts 41% on Terminal-Bench 4.0, up thirteen points from the public preview, and 48% on the Artificial Analysis Coding Agent Index, effectively tied with GLM-5.3 while using 47% fewer tokens per task. Elsewhere: 19.7% on Cognition FrontierCode 2.0, 51% on the AA Cyber Index, 73.3% on FinWorkBench and 84.5% on BrowseComp.
The licence is where to look before you plan around it. Mistral describes it as a modified MIT and says the model is free to use, modify, fine-tune and commercialise — but businesses that operate a Model-as-a-Service offering must obtain a separate commercial licence from Mistral. If you are reselling inference, this is not MIT.
One figure to note: Mistral’s own announcement says 49B active parameters, while the Hugging Face repository is named Mistral-Large-4-1T-A52B. Those do not agree, and Mistral has not reconciled them.
Sources: Mistral AI · Hugging Face.
Published 8 November 2026. Every figure above was checked against the reporting linked here on the same day.