In September 2026 I shared a series of LinkedIn updates covering my move back to the office in San Francisco, progress on MiMo‑V2.6‑Pro, Xiaomi’s Groupwise Advantage Redistribution, the fly‑brain chess model, and a strategic compute‑allocation proposal.
MiMo‑V2.6‑Pro’s run cost $2.6 M and took only 30 training steps over 123 hours, each step rolling out 25 K trajectories and processing 2.7 B‑3.7 B tokens.
The MoE router was frozen for the entire run to prevent expert‑balance collapse.
Xiaomi’s Groupwise Advantage Redistribution prevents rapid growth in turns and token length while allowing pass rate to keep improving.
The fly‑brain chess model contains 138,639 neurons and 15,091,983 connections.
It recovers Stockfish’s top move 30.4 % of the time on 4.4 M annotated Lichess positions.
I allocated 99 % of compute to research and 1 % to inference, ensuring ample training capacity while outsourcing inference once models are strong.
I also advocated for an open‑weight release strategy to foster external evaluation and rapid learning loops.