🤖Mistral Large 4 Preview: 1T Parameters, 49B Active
TL;DR
Mistral's Large 4 preview is a 1 trillion parameter model with 49B active, trained on 3,800 NVIDIA Grace Blackwell GPUs. Simon Willison's verdict: not frontier class, but Mistral is back to roughly six months behind.
Mistral's Large 4 preview is a 1 trillion parameter model with 49B active, trained on 3,800 NVIDIA Grace Blackwell GPUs. Simon Willison's verdict: not frontier class, but Mistral is back to roughly six months behind.
Key Points
1T total and 49B active parameters, available as an API preview now
Open weights promised for the end of October
Only 'none' and 'high' reasoning levels are exposed through the API
Scores 38 on Artificial Analysis, just behind DeepSeek 4.1 Flash (552B); Mistral Large 3 scored 9
Simon's pelican test: 'high' used 2,717 output tokens versus 3,275 for 'none', with a better drawing
Why It Matters
A jump from 9 to 38 in one generation puts a European open-weight lab back in the conversation. The catch is that the weights haven't shipped, so treat the benchmark as a promise.
Quick Facts
Comments
Be the first to comment
Enjoyed this article?
Get it daily. 7am. Free. Reads in 5 minutes.
Join 3,564 builders reading daily.