Mistral Releases Weights for 1-Trillion-Parameter Large 4, Halves API Prices
Mistral has released the weights for its 1-trillion-parameter model, Mistral Large 4, while cutting API prices by approximately 50%. The model, refined through additional reinforcement learning, outperforms GLM-5.3 on Terminal-Bench but ties it on the Coding Agent Index, with a key advantage in reduced token usage per task. Notably, the licensing terms have shifted from Apache 2.0 to a modified MIT license, imposing stricter commercial requirements on Model-as-a-Service providers.
Full Report
Rohan Paul (@rohanpaul_ai) · 2026-10-29 18
Mistral released the weights of its 1-trillion-parameter Mistral Large 4 and roughly halved its API price.
Unlike Large 3’s Apache 2.0 release, the weights carry a modified MIT license that makes Model-as-a-Service providers obtain a separate commercial license.
says 3 more weeks of RL since the preview lifted the model’s Terminal-Bench 4.0 score from 28% to 41%, 1 point above China’s GLM-5.3.
On Artificial Analysis’s Coding Agent Index, Large 4 does not lead: its 48% effectively ties GLM-5.3, and its edge is 47% fewer tokens per task.
API pricing is now $0.70 input, $0.07 cached input and $2.10 output per million tokens, about half the list prices of $1.36, $0.14 and $4.18 that applied by 22 October with no sale running.
The Mistral Ecosystem Open-Weight license lets anyone use, modify and commercialize the model but requires Model-as-a-Service businesses to get a separate commercial license, a tighter deal than the Apache 2.0 license on Mistral Large 3’s open weights.
