Mistral Large 4 has 1 trillion parameters and only 5% of them work on each token. Here is the server it takes to run.
Mistral's new model, in preview since 6 October 2026: about 52 billion active parameters, a 1 million token context, $1.36 in and $4.18 out per million tokens, and about 1,050 GB of weights at FP8.