Xiaomi Corp. (1810.HK): Our takeaways from MiMo-V2.6 RL live streaming

Xiaomi MiMo started public live streaming of its MiMo-V2.6 RL (Reinforcement Learning) process 4 days ago, which stopped post 30 steps of RL for both V2.6-Flash and V2.6-Pro, marking the first time a major AI lab has exposed live post-training telemetry in real time. RL training has led to substantial improvement in long-horizon agentic performance: likely using a similar base to MiMo-V2.5 series, V2.6-Flash/Pro achieved DeepSWE scores of up to 67.86/72.57 (avg@3) within 30 steps of training. MiMo-V2.6-Flash appears to provide more favorable RL investment returns, achieving 19+ DeepSWE score gains within 3.5 days on US$0.85mn total spend, compared to V2.6-Pro which cost US$2.62mn on 14+ score gains. Mid-stream interventions revealed by live streaming (infra error re-runs, bad-pattern datasets, zero-gradient tasks, GPU VRAM out-of-memory) imply RL bottlenecks have shifted to engineering optimization. Apart from the upcoming MiMo-V2.6 series release, Xiaomi continued to advance its AI ec