Xiaomi livestreams MiMo-V2.6 RL training, burning over $1.08M in 36 hours
量子位 · wechat · 2026-09-17
Xiaomi is doing something rare in the industry: livestreaming the full RL training of MiMo-V2.6's Flash and Pro models, with costs, steps, reward curves and GPU failures all public — over $1.08M burned in 36 hours ($30K/hour). Lead Luo Fuli says the past six months went into one question: how far can RL scale? Xiaomi's answer has three axes: scaling training compute (2B tokens per step, 1568 prompts × 16 rollouts, fully async pipeline), scaling environments and harnesses (code, general tasks, vision, chat mixed in one run), and scaling grader compute with in-group credit assignment. Still early: Pro's dynsam/avg@n rose from 0.565 to 0.614, Flash from 0.514 to 0.603; DeepSWEv1.1 offline evals are 62.24 and 60.77 respectively.
Related event: Xiaomi live-streams MiMo-V2.6 RL training, burning over $1.08M in 36 hours(14 posts)→
More from Models
- Anthropic's server-side rule reignites the case for owning open models over renting AI — alex_verem · 2026-11-02
- Epoch AI: open-weight GLM 5.3 exploited the patch gap to hack 6 Firefox releases at ~$1,000 each — rohanpaul_ai · 2026-11-02
- SemiAnalysis: safety classifiers sandbag scores of Opus 5.5 and GPT-6 Astra — morqon · 2026-11-02
- Liquid AI's 1.7B LFM3 becomes first local model to nail live fixtures lookup on a MacBook — NielsRogge · 2026-11-02
- Claude upgrade-page traffic rebounded to 17.6M in September, near May's 18.6M peak — firstadopter · 2026-11-02
- Open-source GLiNER-PII scrubs 13 identifiers from a discharge note in 550ms on GPU — MaziyarPanahi · 2026-11-02