GitHub radar
Mistral releases its largest AI model, a trillion params
French AI company Mistral released Mistral Large 4, its biggest model yet — open weights, but out of reach for home hardware.
French company Mistral released Mistral Large 4, its largest model so far. The whole network has about a trillion parameters, but answering one word only turns on a part of it: around 49 billion active parameters, while the rest are expert blocks that stay idle until needed. The model reads text and images together, and it keeps up to a million tokens in memory in one conversation, which is hundreds of pages of text. Mistral trained it from scratch on 3,800 Blackwell accelerators in its own datacenters in Europe. On a coding-agent benchmark it scores 48%, about the same as the open Chinese model GLM-5.3, but it uses 47% fewer tokens for the same task, so the result costs less to get. On a broader intelligence benchmark this is the best score among western open-weight models, but Chinese open models still score higher on that same index.
Why a vibe-coder should care
In Claude Code this model changes nothing for me, I work in Opus and Fable, and an open model of this size is not something you install at home anyway. But as a reference point it is honest: it shows that the strongest open model out of Europe still trails Chinese open models on general intelligence, even while it holds its own on coding and costs less per task there. If you happen to run your own server with several strong GPUs, this is one of the rare cases where a trillion-parameter model can actually be deployed by you, and not only called through someone else's API.
How to install
Copy this and send it to your agent — Claude Code, Codex, any of them:
Open the Mistral Large 4 model card on Hugging Face: https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B — explain what it's made of, what hardware it needs to run, and how it differs from earlier Mistral models.
A server-only model — not something you run at home. It was trained on 3,800 Blackwell-class GPUs, and even the company's own deployment examples call for a multi-GPU cluster of the same class, not a single machine.
Open on Hugging Face▌ More finds