Mistral releases the weights of its trillion-parameter model

Under the licence, anyone may run the model, adapt it and build paid products on it. The one exception is a firm offering the model itself as a hosted service, which must first get a separate commercial licence from Mistral. Those who would rather not host it can call Mistral’s API at $0.70 per million input tokens and $2.10 per million output tokens, or use it through Microsoft Foundry.

Mistral says three more weeks of training made it the best open model for coding. On Artificial Analysis’ Coding Agent Index it scores 48 to the 49 of China’s GLM-5.3, ZDNET reports, while on the Terminal-Bench 4.0 coding benchmark it edges ahead, 41 per cent to 40. Mistral also says it uses 47 per cent fewer tokens per task than GLM-5.3, TechCrunch notes.

Our cloud GPU price table, read on 24 October 2026, lists the H100 with 80 GB of memory at $1.99 an hour at the cheapest.

At 8 bits a parameter, a common way to serve large models, Le Chonk’s weights alone fill about 1,000 GB, the memory of at least 13 of those cards before any working memory, roughly $26 an hour at that rate, and more in practice, since 13 cards span two eight-GPU servers. For a brand writing ad copy, the per-token API is the realistic route, and the download matters to firms that already run GPU clusters.