XNews

MiMo V2.6 Pro open-source model with large-scale RL

MiMo V2.6 Pro open-source model with large-scale RL

Xiaomi released MiMo-V2.6 Pro and Flash, open-source multimodal models. Pro scored 46 on Artificial Analysis, tying with Grok 4.7 as the strongest open-source model. It used large-scale reinforcement learning (30 steps, 75k trajectories, $2.62M cost) with a MixRL strategy. API pricing starts at $0.435 per million input tokens.

Key insights

MiMo-V2.6 Pro achieves top open-source model score

The Pro model scored 46 on the Artificial Analysis comprehensive intelligence index, tying with xAI's Grok 4.7 and surpassing Kimi K3 and Qwen3.8 Max. It remains behind closed-source leaders Claude Fable 5.1 and GPT-6 Astra. The model supports 1M token context and multimodal input (text, image, audio, video).

Training used largest single RL run for an open-source model

The training cost $2.62 million and covered 30 steps with about 750,000 trajectories, each averaging 110,000–150,000 tokens. Team lead Fuli Luo, who previously worked on DeepSeek R1, said the engineering challenges exceeded R1. The MixRL strategy combined programming, vision, and cybersecurity tasks in one RL run, with separate training for verification-hard tasks merged via MOPD.

Aggressive pricing: 1/20 to 1/60 of overseas models

Pro API pricing is $0.435 per million input tokens and $0.87 per million output tokens; Flash is $0.14 and $0.28 respectively. Xiaomi claims the price is 1/20 to 1/60 of comparable intelligence from overseas models. Flash scores only slightly below Pro on many agent benchmarks, offering high cost-effectiveness for bulk agent tasks.

Beyond model weights: open-sourced RL framework and 7,000 tasks

Xiaomi released the complete RL training framework, over 7,000 RL task environments, and a distilled Qwen-9B model from MiMo RL trajectories. This enables other teams to replicate large-scale RL training on smaller models without building a trillion-parameter base model from scratch. The models are released under MIT license.

Source posts