Wall-OSS

Wall-OSS is an open-source, end-to-end embodied AI foundation model from Independent Variable Robotics (自变量机器人), released alongside training code in September 2025 as the company's roughly 1 billion yuan Series A+ round was announced . Architecturally, it pairs a "shared attention + expert FFN split" design with a three-stage training paradigm ("discrete first, then continuous, then joint") to carry a vision-language model's cognition over to physical manipulation without loss, and it claims stronger generalization on long-horizon tasks than other foundation models, per a 九合创投 piece . CTO Wang Hao has framed the open-sourcing as a bet on data-driven end-to-end training — "language, vision, and action should all be represented and aligned in the same space," with a unified architecture at training time and flexible cloud/edge splits at deployment . The company has since iterated the line with WALL-OSS-0.5 and, by mid-2026, the WALL-B and WALL-WM world models built on related foundations .