Archive — CN — technode
Xiaomi open-sources MiMo-V2.6 models after scaling reinforcement learning
Xiaomi has open-sourced its MiMo-V2.6 series, including top models MiMo-V2.6-Pro and MiMo-V2.6-Flash, which excelled in reinforcement learning benchmarks. MiMo-V2.6-Pro achieved a score of 46 on the Artificial Intelligence Index, outperforming several competitors. The models provide enhanced capabilities like 3D spatial reasoning and multimodal perception, after completing extensive training quickly and economically.
Key points
- Xiaomi released MiMo-V2.6 models, including MiMo-V2.6-Pro and MiMo-V2.6-Flash.
- MiMo-V2.6-Pro achieved a score of 46 on the Artificial Intelligence Index.
- It outperformed Kimi K3 and GLM-5.3 models which scored 44 and 45 respectively.
- The highest closed-source model scored 53.
- The models completed 30 reinforcement-learning steps in under six days with about 750,000 training trajectories.
- The costs reported were $2.62 million for Pro and $850,000 for Flash.
- New capabilities include 3D spatial reasoning and multimodal perception.
Why it matters
- The development aligns with increasing trends in AI model openness and accessibility as companies look to innovate through collaboration.
- Xiaomi's focus on multimodal AI models reflects the growing trend towards integrating AI capabilities across various mediums and applications by 2025.
Xiaomi's open-sourcing of the MiMo-V2.6 models marks a significant advancement in AI accessibility and performance, setting a competitive standard in the rapidly evolving AI landscape.