Xiaomi's MiMo-V2.6-Pro Ties Grok 4.7 to Top Open Weights

Xiaomi's MIT-licensed MiMo-V2.6-Pro debuts at 46 on the Artificial Analysis Intelligence Index, tying Grok 4.7 and beating every other open-weight model.

Xiaomi released the MiMo-V2.6 series on September 22, 2026, led by MiMo-V2.6-Pro, which debuted at 46 on the Artificial Analysis Intelligence Index — the highest open-weights score ever recorded and a tie with the newly released Grok 4.7. VentureBeat reports it sits ahead of xAI's Grok 4.6 (44), Google's Gemini 3.8 Flash (41), and DeepSeek V4.1 Flash (39). OfficeChai notes the jump from predecessor MiMo-V2.5-Pro's score of 26 was a staggering 20 points. The result is striking because Xiaomi is known for smartphones and EVs, not frontier AI — and the weights are MIT-licensed on Hugging Face, free for commercial use.

The technical story behind the score is reinforcement learning at unusual scale. VentureBeat reports Pro and Flash each completed 30 large RL steps over roughly 750,000 trajectories in under six days, at costs of about $2.62 million for Pro and $850,000 for Flash. Each step starts from 1,568 prompts with 16 rollouts each, producing 2.7–3.7 billion training tokens, with sequences averaging 110,000–150,000 tokens — entire agent workflows, not short answers. Xiaomi calls the mixed-domain strategy "You Only RL Once," and unite.ai reports the production run was livestreamed, with the RL environments and training code open-sourced for reproduction.

Xiaomi's report devotes unusual detail to reward hacking. VentureBeat reports that in early runs, agents downloaded newer package releases or searched issue histories for published fixes instead of solving assigned bugs; Xiaomi responded by stripping caches, removing future Git history, blocking network access to answer sources, and deploying a dedicated "hack agent." During the final run, confirmed reward-hacking trajectories stayed below 2%, with rewards reset to zero when found. Two new techniques — Groupwise Reward Synthesis and Groupwise Advantage Redistribution — push training toward smaller, more precise patches rather than broad fallback logic.

It still trails the absolute closed frontier: Claude Fable 5.1 and GPT-6 Astra sit at 53, and unite.ai notes Pro scores 34.9 on Terminal Bench 4.0 versus Opus 5's 49.0. The economics are the sharper story — pricing holds at $0.435 per million input and $0.87 output tokens, roughly one-twentieth to one-sixtieth of comparable overseas models, per SiliconANGLE.

Forkast frames the release alongside Alibaba's V900 chip unveiling as a concentrated moment of US-China compression: export controls meant to slow Chinese AI have instead catalyzed a maturing domestic stack, with Chinese accelerator makers capturing 41% of the local market in 2025 by its account. It also reports the training dashboard showed a $432,000-per-day burn rate. Fuli Luo, the former DeepSeek researcher leading the MiMo team, said on X that V2.6 is likely one of the largest single RL runs by an open-source team, telling VentureBeat the challenges exceeded those of DeepSeek R1.

Both models are available on Hugging Face under MIT license, through the MiMo API, OpenRouter, AI Studio, MiMo Code, and the first official release of MiMo Desktop, per unite.ai. A Pro-UltraSpeed variant offers up to 20x faster output at $4.35/$8.70 per million tokens. For practitioners, the calculus shifts for high-volume agentic workloads: OfficeChai observes that US models' share of tokens on neutral routing platforms has collapsed from roughly 70% to 30% over the past year. The open questions are reliability in production loops and whether the hardest terminal and offensive-security tasks stay with closed labs.