Light-O1-Preview

Light-O1: Scaling Whole-Body Intelligence with Human Action Pretraining

Light-O1-Preview is a reasoning text-to-action model: given an instruction in natural language it reasons about intent and constraints, then generates a sequence of whole-body humanoid actions. Reasoning text and actions are decoded in one autoregressive stream — action codes are ordinary rows of the embedding table rather than a task-specific prediction head — so the model's reading of an instruction can be inspected directly instead of inferred from the resulting motion. Actions are produced in a shared humanoid representation rather than for one specific robot, so a generation can be retargeted to other embodiments.

Inference, serving and the operator console: Light-O1 repository. Try it without any local setup in the Light-O1-Preview Playground.

Files

Path	Content
`config.json`, `model.safetensors`	model weights
`tokenizer.json`, `tokenizer_config.json`, `chat_template.jinja`, `processor_config.json`, `generation_config.json`	tokenizer / processor
`human_action_138_v1.json`	the action representation that must match training (field layout, 22-joint skeleton, frame rate, axis convention); read automatically by the inference code
`action_tokenizer/`	FSQ action-decoder bundle (codebook 65,536, four levels of 16, 20 fps)

Usage

hf download LightOriginsHQ/Light-O1-Preview --local-dir ./Light-O1-Preview

git clone https://github.com/lightorigins/Light-O1 && cd Light-O1
uv sync --extra inference

# one-off generation on the command line
uv run --extra inference light-deploy --model /absolute/path/to/Light-O1-Preview \
    --prompt "a person waves with the right hand" --thinking --output human_action.npy

# local WebUI: streamed reasoning and a 3D preview of the generated action
uv run --extra inference light-deploy-server \
    --model-path /absolute/path/to/Light-O1-Preview --port 8090

GPU inference needs Linux x86-64, Python 3.11 and a CUDA 13 compatible environment. The HTTP API, the split GPU / Control host setup and the optional MuJoCo simulation example are documented in the code repository.

Model details

Input: a natural-language prompt.
Output: a reasoning trace followed by action codes, decoded to human_action_138_v1 — per-frame root displacement in the ground plane, pelvis height, yaw rate, 22 parent-local 6D joint rotations and two hand-open scalars, right-handed and Y-up at 20 fps.