Mistral publishes the weights of its 1-trillion-parameter Large 4
The benchmark results are Mistral's own; no independent rerun was found. The licence is not a standard open-source licence, and hosting the model as a paid service needs Mistral's permission. API pricing has not been published, and running the full model still needs data-centre hardware.
What was claimed?
Mistral published the weights of Mistral Large 4 on Hugging Face, three weeks after an API preview. The model card describes a 1-trillion-parameter mixture-of-experts model that reads text and images, with about 52 billion parameters active per token, trained on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European data centres. It is released under the Mistral Ecosystem Open-Weight License, a modified MIT licence that allows commercial and non-commercial use but requires a separate commercial licence to run a model-as-a-service business. TNW reported on October 29 that Mistral says further reinforcement learning raised its Terminal-Bench 4.0 score from 28% at preview to 41%.
What does the evidence support?
Company claim, independent confirmation unresolved
Mistral calls it its largest and most capable model to date, and downloadable weights let governments and companies run a trillion-parameter model from a European lab on their own hardware rather than depend on a provider's service. Mistral says a quantised version runs on four NVIDIA B200 or B300 GPUs.
What are the limits?
The benchmark results are Mistral's own; no independent rerun was found. The licence is not a standard open-source licence, and hosting the model as a paid service needs Mistral's permission. API pricing has not been published, and running the full model still needs data-centre hardware.
Claim or evidence date: 2026-10-29. Last verified: .
Colors classify evidence; they do not rank truthfulness. Explore hype versus reality →