Six days, $3.5 million — Xiaomi livestreamed an entire reinforcement-learning run for MiMo-V2.6, with a live dollar meter spinning through the training. Pro finished 30 steps in 5d3h for ~$2.6M; Flash in 3d10h for ~$0.9M.

What the money bought

MiMo-V2.6-Pro packs 1.02T total parameters with 42B active, scoring 46 on the Artificial Analysis Intelligence Index — No.1 among open-source models globally — and approaching Claude Opus 5 and GPT-5.6 Sol on most agent benchmarks; on AutomationBench it scored 53.1, ahead of both. On the out-of-sample DeepSWE v1.1, Pro jumped from 58.4 to 72.57 and Flash from 48.7 to 65.68 — real generalization, not leaderboard cramming.

Pricing stayed aggressive: Pro at ~RMB 3/6 per million tokens in/out, with an average task cost of $0.13 on AA's measure — 29x cheaper than the closed-source Grok 4.7 xHigh ($3.74) released the same day. UltraSpeed mode reaches ~900 TPS.

The full open-source bundle

Weights, the complete technical report, 7,000+ RL environments, an end-to-end RL training framework, composable mini-harnesses — all released, plus a Qwen3.5-9B-based distill SFT'd on 77.4B tokens. Hugging Face's CEO publicly praised the drop; Ai2's Nathan Lambert hailed the "top open model + live RL dashboard" combo.

From Vibe Coding to Vibe World

Right after training, Xiaomi's advanced materials team threw Pro at an unrehearsed science task: designing MOFs to capture PFAS "forever pollutants". The model searched literature, formed hypotheses, checked novelty, built simulations and ran open-source tools — yielding two UiO-67-type zirconium candidates with million-to-ten-million-fold adsorption gains. A PKU professor rated its performance "that of a well-trained doctoral researcher"; the internal R&D cycle shrank from a month to 2-3 days.

For China's AI industry, the livestream's significance goes beyond Xiaomi: front-tier RL training's cost, process and output are now on the table for everyone — and the open/closed gap is closing at "transparency + price-performance" speed.

(Facts aggregated from public reporting)