China’s Orca world model reportedly matches the performance of specialized robotics systems without ever training on a single action label (The Decoder). The result is notable for physical-AI work because it points toward learning control from unlabeled interaction rather than the curated action datasets such systems have historically required.
Separately, Ant Group’s Robbyant unveiled LingBot-VA 2.0, a causal video-action model built natively for physical AI (MarkTechPost), continuing a run of embodied-AI model releases from Chinese labs.
In coding models, Meta’s Muse Spark 1.1 outperforms Z.ai’s GLM-5.2 on coding tasks while costing slightly less, according to reporting (The Decoder), positioning Meta’s new entrant directly against the open-weight coding models that have gained traction with developers.