Xiaomi Reveals Training Details of 1T Model MiMo-V2-Pro: Thousands of GPUs Used, No Job Levels or Deadlines

According to monitoring by Dongcha Beating, Luo Fuli, head of Xiaomi’s large model team, disclosed in her first in-depth interview that the MiMo-V2-Pro model base has a total parameter count of 1T, utilizing thousands of GPUs for training. She believes that a scale of 1T is the baseline for achieving performance close to Claude Opus 4.6 and securing entry into the next phase of agent competition. On a technical level, the Pro version pushes the ratio of global attention to sliding window attention to an extreme sparse ratio of 7:1, controlling the reasoning cost for long texts while expanding the parameter count, and continues to use the MTP (Multi-Token Prediction) architecture to leverage excess computing power for accelerated inference. On the management side, only about 30 to 40 out of the hundred-member MiMo team are directly involved in core iterations, with no established job levels, clear group divisions, or delivery deadlines. When encountering unstable numerical issues such as sudden changes in training loss, the team opts to halt training for troubleshooting, even if it means stopping for one or two weeks and incurring millions in computing costs.