MiMo-V2.6 scales compute and environments in active RL run

Luo Fuli: Luo Fuli leads the MiMo AI team at Xiaomi and previously worked as a researcher at DeepSeek. He oversees the development of the MiMo model family, including post-training techniques such as reinforcement learning. His recent work centers on exploring the limits of RL scaling for AI self-improvement, with public updates on ongoing training runs. MiMo-V2.6: MiMo-V2.6 is an AI language model under development by Xiaomi's MiMo team, focused on advancing reasoning and agentic capabilities through reinforcement learning. It represents the next iteration in the MiMo series of models designed for complex tasks including coding, multi-step planning, and autonomous agent workflows. The model is currently in an active RL training phase emphasizing scalability in compute, environments, and evaluation methods. Training Focus: The MiMo team dedicated nearly half a year to studying the scalability of reinforcement learning as a path to AI self-improvement. Team Background: The effort builds on prior MiMo releases emphasizing agentic intelligence and multimodal capabilities from Xiaomi's dedicated AI research group. Public Streaming: The ongoing RL training run for MiMo-V2.6 is being streamed publicly, with plans to release technical details incrementally.