XFEON unveils Physical-AI matrix: Xinghe S1 edge chip, AGLobe brain, Xinghui Token workstations

Chengdu-based XFEON on September 22, 2026 launched a Physical-AI stack spanning the Xinghe S1 edge chip (~105 TOPS INT8 at ~25 W, over 4 TOPS/W), the AGLobe embodied brain built on concept learning, and Xinghui Token workstations for private deployment of models up to 671B parameters.

• Xinghe S1: ~105 TOPS INT8 per die, ~25 W, efficiency above 4 TOPS/W
• Claims ~50% higher inference throughput and ~85% lower model memory
• Xinghe R (up to ~420 TOPS), K for satellites and E for edge/industrial
• An eight-card Xinghui node can host a 671B-parameter model