Mistral Large 4 previews one trillion parameters

What's this about?

Mistral's natively multimodal mixture-of-experts model has one trillion total parameters but activates 49B for each token. API access is available now; weights are promised by the end of October.

What exactly is an active parameter?

A mixture-of-experts model routes each token through only some specialised subnetworks. Total capacity can be huge while compute per token reflects the smaller active subset.

What's the community saying?

1. The 49B-active design makes self-hosting more plausible than the 1T headline suggests, though still demanding.
2. European training and deployment appealed to sovereignty-sensitive customers.
3. Users want independent tests and the actual weights before accepting “best open model” claims.