Xiaomi's MiMo-V2.6-Pro Takes the Open-Weights Crown — Tying Grok 4.7 on a Trillion-Parameter MoE Trained on Chinese Chips
On September 21, 2026 — the same day SpaceXAI shipped Grok 4.7 — Xiaomi released MiMo-V2.6-Pro, and per VentureBeat it "has arrived as the top-performing open-weight model in the world on third-party benchmarking firm Artificial Analysis' Intelligence Index, scoring 46." That ties Grok 4.7 (46), and puts it ahead of closed proprietary models including Grok 4.6 (44) and Gemini 3.8 Flash (41) — and above fellow Chinese open-weights models DeepSeek V4.1-Flash (39) and V4.1-Pro (36).
The economics are the disruptive part. The model is MIT-licensed with downloadable weights on Hugging Face, priced at $0.435/$0.87 per million input/output tokens — a fraction of Grok 4.7's $2/$6 and roughly 1/70th of Claude Fable 5.1's $10/$50. The smaller MiMo-V2.6-Flash ($0.14/$0.28) stays within a few points of Pro on most agent benchmarks. Both have 1M-token context and native multimodal input.
How they did it — RL, not scale. Xiaomi ran 30 large RL steps over 750,000 trajectories in under six days, at reported costs of **$2.62M for Pro and $850K for Flash**, using a mixed-domain "You Only RL Once" strategy with elaborate reward-hacking defenses (confirmed hacking trajectories held below 2%). Fuli Luo, the ex-DeepSeek researcher leading MiMo, wrote: "In an era when compute is brutally scarce, we still chose to dedicate a team of several dozen people to one goal over an extended period: scaling up RL." VentureBeat also quotes Paper Instruments CEO Daanish Khazi arguing the release should make researchers "update every prior on post-training scaling laws," noting the capability was extracted "entirely on Chinese chips rather than Nvidia GPUs."
Unusually, Xiaomi open-sourced more than weights: the technical report, 7,000+ RL task environments, an end-to-end RL framework, composable mini-harnesses, and a distilled 9B model — a potential reproduction blueprint for agent-RL without a trillion-parameter budget.
Caveats: Xiaomi's own benchmark suite shows Claude Opus 5 still ahead on DeepSWE v1.1, ProgramBench, and Terminal Bench 4.0, and GPT-5.6 Sol ahead on some cybersecurity evals; many agent evals are vendor-run. This is not a benchmark sweep — it is the open-weights frontier moving to parity with the newest closed releases on composite intelligence per dollar. It lands two days before the Trump-Xi summit (see US–China: From "No AI Guardrails" to an AI Incident Channel and a November Dialogue) and alongside Huawei's domestic-silicon push (Huawei Doubles Down on Domestic AI Silicon: Two Ascend 960 Chips Slated for 2027 Despite US Curbs).