Xiaomi open-sourced a trillion-parameter model trained in one RL run

Xiaomi's MiMo-V2.6-Pro-RL is a 1.02T open MoE with 42B active and 1M context; Xiaomi claims parity with Claude Opus 5 on agent tests. Open weights now reach the trillion scale US labs keep closed. Run its eval suite on your own tasks this week.