Light-O1-Preview

Light-O1-Preview is an open-weight text-to-action model from Light Origins, an embodied-AI company with offices in Beijing, Shenzhen and Singapore. Given a natural-language instruction, it first writes a short reasoning trace about what the body needs to do and then generates a sequence of whole-body humanoid actions. Light Origins introduced it in a technical blog post dated 21 September 2026 as the public preview of Light-O1, its line of embodied foundation models; the Hugging Face repository was created on 20 September.

The model is built on Qwen3.5-4B, and Hugging Face's weight metadata shows about 5.5 billion parameters in BF16; press reports describe it as a 6B model. Actions are produced in a shared humanoid representation at 20 frames per second, covering root motion, pelvis height, yaw rate, 22 joint rotations and two hand-open values, so a generation can be retargeted to other robots. Light Origins says the wider Light-O1 approach pretrains on human actions recovered from internet video and then adapts to specific robots. Inference needs Linux, Python 3.11 and a CUDA 13-compatible NVIDIA GPU; a browser playground is available on Hugging Face Spaces. The weights and code are released under Apache 2.0.

In its own evaluation on HY-Motion-Bench, Light Origins reports an overall score of 78.0, against 74.7 for HY-Motion-1.0 and 61.4 for Kimodo.

This is a research preview, not a robot controller. Driving a physical robot requires a separate behaviour foundation model or low-level controller, and the released Unitree G1 example runs in MuJoCo simulation. RuntimeWire notes that the preview is narrower than the full Light-O1 system shown in the company's demonstrations, so buyers should not expect the preview alone to reproduce them.

Not yet independently verified: Evaluation results are Light Origins' own. The parameter count comes from Hugging Face's weight metadata (about 5.5B); the developer does not state one, and press reports round it to 6B. We will update this when it can be confirmed, and remove this note.