Modelbest Open-Sources MiniCPM-Robot Embodied Models

量子位 · wechat · 2026-07-20

During WAIC, Modelbest announced and open-sourced the embodied AI model series MiniCPM-Robot and the inference framework PhyAI. Key releases include:

MiniCPM-RobotManip: A 1.5B-parameter general VLA model for robotic arm control. By incorporating historical observations and action memory, it significantly outperforms π0.5 on the in-context memory benchmark RMBench. Combined with visual Token compression and streaming computation, its single-decision latency on an H100 is about 120 milliseconds, nearly half that of π0.5.
MiniCPM-RobotTrack: A 0.9B-parameter tracking and navigation model deployed locally on robot dogs. It maintains single-target tracking, multi-person interference tracking, and fuzzy target tracking even in weak or no network environments, achieving the best tracking rates among open-source solutions.
PhyAI Framework: Co-developed by Mingti Technology and academic institutions to solve adaptation and acceleration challenges when deploying embodied models to physical hardware, delivering significant speedups for mainstream VLA models across various GPUs.

Additionally, Modelbest has partnered with Leju and Geely to integrate edge multimodal large models with robotic hardware. This will be deployed in real-world scenarios like exhibition tours, park patrols, and production warehousing, supporting offline operation and data loop governance.

Original post →

Modelbest Open-Sources MiniCPM-Robot Embodied Models

More from Embodied