Xiaomi's livestreamed RL post-training cost $3.47M, lifts MiMo-V2.6 to top open model
AlchainHust · x · 2026-09-22
Xiaomi livestreamed its post-training, then shipped MiMo-V2.6
- Xiaomi ran a public livestream of MiMo-V2.6's post-training: steps, eval curves and costs all visible in real time ($2M burned in two days). An AI2 researcher (author of the RLHF book) called it the coolest public at-scale RL material available.
- Final bill: $3.47M — Flash cost $850K, Pro $2.62M, 30 RL steps each. Per Luo Fuli this may be one of the largest single RL runs by compute ever done by an open-source team.
Results
- MiMo-V2.6-Pro tops Artificial Analysis' open-model rankings with an intelligence index of 46, at roughly $0.13 per task.
- DeepSWE v1.1: Pro 58.4 → 72.6, Flash 48.8 → 65.7 — same base model, gains entirely from post-training.
- RL code, training environments and tech report are open-sourced for reproduction (if you have $3M+).
More from Infra
- Show Reddit: ge, a Rust GGUF engine, benchmarks against llama.cpp on a 4GB RTX 3050 — Raptor2496 · 2026-11-12
- Musk: 100GW of space solar equals ~800GW on Earth, and ground solar needs batteries — elonmusk · 2026-11-12
- Real-world test: Kimi K3.1 Pro costs 5.4x less than Opus 5.5 per task despite 2x list-price gap — Arindam_1729 · 2026-11-12
- Dino10B: first 10-billion-scale nearest-neighbor search benchmark, 10.3TB of DINOv3 embeddings — philipvollet · 2026-11-12
- Many announced data centers may never get built, TechCrunch argues — TechCrunch AI · 2026-11-12
- Meta loses longtime infrastructure head Santosh Janardhan after 16 years — TechCrunch AI · 2026-11-12