#xiaomi

The LocalLLaMA angle is not just the 1000+ tps headline, but whether FP4, DFlash, and commodity GPU kernels can be reproduced outside Xiaomi’s hosted trial.

LocalLLaMA lit up because Xiaomi MiMo dropped an MIT-licensed MoE with 1.02T total parameters, 42B active parameters, and a 1M-token context window. The excitement was real, but so was the hardware reality check: people loved the openness and agentic claims while joking about how many serious GPUs you still need.

This matters because Xiaomi just put a frontier-scale model family behind permissive terms instead of a closed API gate. The MiMo-V2.5 release promises a 1M-token context window, MIT licensing for commercial use and fine-tuning, and a Pro variant Xiaomi says leads open models on GDPVal-AA and ClawEval.

Xiaomi's humanoid robot completed three consecutive hours of autonomous operation at a Beijing EV factory's die-casting workshop, achieving a 90.2% success rate in nut installation tasks. Powered by a 4.7B-parameter in-house VLA model, it meets the production line's 76-second cycle time requirement.