Light Origins has made Light-O1-Preview, a model for full-body actions containing approximately 6 billion parameters, publicly available. This model is released under the Apache 2.0 license on Hugging Face and GitHub, according to the company's technical blog from September 21st and its English coverage.
This preliminary version of the model represents a public segment of Light-O1—Light Origins' foundational line of embodied systems. This line is first trained on structured human actions extracted from video material and then adapted to various robotic bodies. The official company branding is Light Origins; this release differs from the earlier LightNav-0 navigation module.
Light Origins described a workflow that involves segmenting people in videos, reconstructing three-dimensional motion, and tokenizing root trajectory, body pose, and hand state into discrete action tokens. These tokens are interleaved with language and visual data for autoregressive pre-training of the Transformer.
Scaling experiments, covering from 3.75 billion to 120 billion multimodal tokens—which corresponds to approximately 100,000 hours of reconstructed human actions at the largest budget—demonstrate a reduction in next-action loss and full-body pose error after adaptation to egocentric human data, as well as to Unitree G1 teleoperation systems and Light Origins' proprietary LightBot datasets.
Demonstrations mentioned in the launch materials transfer this prior experience to LightBot and Unitree G1 to perform actions such as squatting, balancing, wiping, and picking up objects.
On the Hugging Face page, the preview shows a 6B checkpoint in the Qwen3.5-4B-Base model tree. The model accepts a text instruction, generates a brief reasoning trace, and outputs full-body action sequences at a rate of 20 frames per second. The output format consists of 138 values per frame, covering root movement, hip height, yaw, 22 joints, and hand openness states.
Applying these trajectories to real hardware still requires a separate behavior model or low-level controller. Light Origins documents an example path for Unitree G1, which is currently running in simulation with an additional policy checkpoint. Results from the company's own RoboCasa GR-1 kitchen simulations are also reported, achieving 79.3% success on macro metrics across 24 tasks, although this is not an independent public rating.
For robotics teams, Light-O1-Preview is a loadable motion-priority layer under a permissive license, not a ready-to-use robot operating system. The open weights allow labs to verify claims about pre-training on human video content; however, the quality of local manipulation tasks between different bodies still depends on the adaptation data and controllers, which remain partially proprietary.
