ModelBest Open-Sources Embodied Model Series MiniCPM-Robot at WAIC

At WAIC 2026, ModelBest released and open-sourced MiniCPM-Robot, its first embodied AI model series, extending its edge-side MiniCPM lineup from multimodal into real-world robotics. The series covers two core capabilities — manipulation and tracking/navigation — and explicitly flags long-horizon context memory as a key problem to solve, signaling edge-model makers pushing deeper into embodied intelligence.

Key Details

MiniCPM-Robot currently includes two models: RobotManip, a 1.5B general vision-language-action (VLA) model for robot manipulation built on ModelBest's own multimodal MiniCPM; and RobotTrack, a 0.9B model for object tracking and navigation. Multiple posts note that RobotManip focuses on the "no context memory" problem robots face in long-horizon tasks, and 新智元 reports it has corresponding evaluations on the RMBench benchmark.

Showcase and Deployment

In its official post, ModelBest framed the release under the theme of "large models entering the real world," presenting it alongside commercial progress of its density-law-based edge models and grouping it under the "embodied and in-vehicle intelligence" direction. The company also announced a jointly developed, offline-capable exhibition-hall guide Agent with Leju Robot, indicating MiniCPM-Robot is meant not just as a research release but as an attempt to land in real terminals and offline scenarios.

2026-07-17 ~ 2026-07-19 · 6 related posts

2 near-duplicate retellings: 赛博禅心 · AI寒武纪