Skip to main content

Xiaomi's MiMo-V2.6: The New Open-Source AI Champion

Xiaomi's MiMo-V2.6: The New Open-Source AI Champion

On September 22, Xiaomi released and open-sourced its native multimodal model, MiMo-V2.6 (Pro/Flash), pushing the boundaries of reinforcement learning (RL) and exploring recursive self-improvement (RSI). The result? A model that's turning heads in the AI community.

Image

Crushing the Competition

MiMo-V2.6-Pro scored 46 points on the Artificial Analysis comprehensive intelligence index, surpassing Kimi K3 and Qwen3.8Max to become the strongest open-source model available. In most Agent benchmarks, it goes head-to-head with Claude Opus5 and GPT-5.6Sol, though it still trails the absolute top closed-source models.

Training That Won't Break the Bank

Both models completed Live RL training in less than 6 days, with costs of approximately $850,000 and $2.62 million, totaling about 750,000 trajectories. The software engineering benchmark DeepSWE v1.1 jumped by about 17 and 14 points, respectively. And they gained new tricks: 3D spatial reasoning and computer operation (CUA) capabilities, enabling tasks like 3D modeling and robotic arm control.

Priced to Disrupt

MiMo-V2.6 keeps the same API pricing as V2.5, making it 1/20 to 1/60 the price of overseas models with similar intelligence. The MiMo Desktop client launched alongside, with an UltraSpeed mode that boosts output speed up to 20 times.

Open Source, Fully Transparent

Xiaomi didn't hold back. They open-sourced model weights, technical reports, 7k+ RL task environments, and an end-to-end training framework. This makes the path and cost of large-scale RL transparent, providing a reproducible foundation for Agentic RL research.

Open-source link: https://huggingface.co/collections/XiaomiMiMo/mimo-v26

Key Points

  • MiMo-V2.6-Pro scores 46 on intelligence index, top open-source model.
  • Trained in under 6 days for ~$850K, with 3D reasoning and CUA capabilities.
  • API pricing is 1/20 to 1/60 of overseas models; UltraSpeed mode 20x faster.
  • Open-sourced weights, reports, 7k+ RL environments, and training framework.
  • Available on Hugging Face: link