Xiaomi’s MiMo-V2.6-Pro becomes top open-weight AI model

Xiaomi, the Chinese electronics and electric vehicle maker, has released MiMo-V2.6-Pro, which has claimed the top spot among open-weight AI models on the Artificial Analysis Intelligence Index with a score of 46 — tying xAI's proprietary Grok 4.7 and leaping past fellow Chinese competitors including Z.ai's GLM-5.3 and Moonshot AI's Kimi K3.venturebeat+2

The result marks a 20-point jump over Xiaomi's previous flagship, MiMo-V2.5-Pro, which scored 26 on the same index. MiMo-V2.6-Pro is MIT-licensed, freely downloadable from Hugging Face, and available through Xiaomi's API at $0.435 per million input tokens and $0.87 per million output tokens. At $0.13 per Intelligence Index task, it costs a fraction of what comparable proprietary models charge — Grok 4.7 costs $3.74 per task for the same score.OfficeChai+2

Xiaomi's own page for the release noted that despite the model's open-weight leadership, "there is still a gap when compared with the strongest closed-source models Claude Fable 5.1 and GPT-6 Astra".mimo.mi

A Cheaper Flash Model for Production Workloads

Alongside Pro, Xiaomi released MiMo-V2.6-Flash, a 310-billion-parameter model with 15 billion active during inference, priced at $0.14 per million input tokens and $0.28 per million output tokens. Flash retains the same 1-million-token context window and multimodal capabilities as Pro while trailing it by only a few points on several agent benchmarks — scoring 67.9 versus Pro's 71.9 on DeepSWE v1.1 and actually outperforming Pro on CyberGym, 95.1 to 94.0.venturebeat

Tim Dettmers, a Carnegie Mellon professor and researcher at the Allen Institute for AI, wrote on X that "the flash model has really good vibes. Feels like the best model in the 300B to 550B class".venturebeat

Reinforcement Learning at Scale

The technical story behind V2.6 centers on what Xiaomi calls one of the largest single reinforcement-learning runs attempted by an open-source team. Each model underwent 30 large RL steps covering roughly 750,000 trajectories in under six days, at reported costs of about $2.62 million for Pro and $850,000 for Flash. Rather than running separate RL programs for coding, visual work and cybersecurity, Xiaomi mixed those domains into a single training run it calls "You Only RL Once."venturebeat

The release arrives at a moment when Chinese open-weight models have established a clear lead over American counterparts. Nathan Lambert, writing in testimony prepared for Congress, noted that Chinese open-weight models are mentioned in over 40% of AI research papers, above the U.S.'s roughly 30%, with the gap widening. American open-weight models trail their Chinese rivals by an estimated 2 to 5 months on capabilities benchmarks.interconnects

Xiaomi is open-sourcing not just the model weights but also more than 7,000 RL task environments, an end-to-end RL framework and composable mini-harnesses — giving outside researchers enough of the stack to reproduce and extend the work.venturebeat