Unitree Robotics has started releasing partial model weights and datasets for UnifoLM-WLA-1.0, its roughly 6-billion-parameter world-language-action foundation model for humanoid robots, after earlier publishing the architecture and 64-task evaluation matrix. Chinese tech coverage dated September 20 says parts of the stack are now downloadable from the project page, while post-training code, UnifoLM-WLA-Base weights, whole-body teleoperation corpora and broader operation datasets remain listed for later drops. A single checkpoint is trained on about 2,500 hours of real-robot data drawn from Unitree open sets and public sources such as BitRobot HIW-500, and is meant to cover 10 whole-body plus 54 tabletop tasks without spinning up a separate policy per chore. The unified action space tokenizes end-effector poses, hand joints and lower-body joints with residual vector quantization so those streams remain time-aligned. Behind the MMDiT action expert sits UnifoLM-ER-1-4B, an embodied-reasoning model built on Qwen3-VL-4B and more than 5 million spatial samples. Unitree reports it leads open models on seven of sixteen multimodal perception benches it tabulated, with figures company-reported, not yet third-party reproduced.