Light-O1 turns internet video into robot motion

Light Origins launched Light-O1, a whole-body intelligence model pretrained on human actions recovered from internet video, with demos retargeted to LightBot and Unitree G1, and released a 6-billion-parameter Light-O1-Preview checkpoint on Hugging Face under Apache 2.0.

HARDWARE desk — robot learning is starved for real motion data; using internet video as a shared pretraining layer before robot-specific fine-tuning is a concrete physical-AI bet you can download and inspect.

The same day the company opened a checkpoint called Light-O1-Preview on Hugging Face, under the Apache 2.0 license. Weights are the learned numbers that make a model work. Apache 2.0 is an open-source license: those weights are public to use and build on, under that license’s own conditions. Hugging Face lists the checkpoint at 6B parameters. A parameter is one learned number. Six billion is the page’s size label, not a count this desk added up. The model card’s own front matter names the starting model as Qwen/Qwen3.5-4B. Qwen is Alibaba’s open model family. The technical post says the scaling runs start from Qwen3.5-4B, which it also calls a 4B base model. RuntimeWire, a same-day wire, says the Hugging Face model tree traces to Qwen3.5-4B-Base. The tree page this desk fetched printed both Qwen/Qwen3.5-4B-Base and Qwen/Qwen3.5-4B. File the card’s own base line as Qwen/Qwen3.5-4B. Do not merge the Base name and the card’s name into one spelling.

What the preview does, still the model card. You type an instruction in ordinary language. The model reasons about what you meant and about the limits of the move, then it writes a sequence of whole-body humanoid actions. The reasoning and the actions come out as one stream. The motion codes sit in the same word list as the language, rather than in a separate motion calculator, so you can read what the model thought the instruction asked for before the body moves. The GitHub page says the same thing in shorter form: the model states in language what the instruction requires of the body, then generates the action. The company also posted a browser playground where you can type a prompt, watch that reasoning, and look at the motion in a 3D view, without installing the model. File the two-part answer as the card’s and the repository’s. This desk did not run a prompt.

The motion uses one shared format, named human_action_138_v1, so the same clip can be aimed at more than one robot body. The card says each frame — one pose in the sequence — stores how the body moved across the floor, how high the pelvis is, how fast the body is turning, the pose of 22 joints, and whether each hand is open. The rate is 20 frames a second. The GitHub page writes the output as a grid of frames by 138 numbers, which is that same layout. File the field list as the model card’s. This desk did not open the weights.

The technical post shows that model on two bodies: Light Origins’ own humanoid, LightBot, and Unitree’s G1. The expressive-skill captions include kneeling on one knee, standing on one leg, wiping a blackboard, and bending down to pick an object up. The GitHub page says that post demonstrates Light-O1-Preview on both robots. RuntimeWire, citing a Light Origins thread on X the same morning, lists kneeling, balancing, wiping a blackboard, and picking up an object, and says the same starting model drives both bodies. File the captions as the post’s. File the thread summary as RuntimeWire’s. A demo on a company video is not a robot you can order today.

The open file is not a finished household robot. The GitHub page says the checkpoint outputs the shared motion, and that a robot still needs a behavior foundation model — a lower controller that turns the intended movement into motor commands. Light Origins uses its own controller for LightBot. For the G1 it names GEAR-SONIC, a low-level motion controller. The example shipped for other people runs inside MuJoCo, a physics simulator, and it needs a separate policy file. The page says that example runs in simulation only. RuntimeWire says the public release is narrower than the household demonstrations: the preview takes text and generates motion, and putting that motion on a robot still takes a separate controller. File that limit as the repository’s and the wire’s.

The scaling study is in the same technical post. Light Origins trained separate models at six budgets, from 3.75 billion to 120 billion multimodal tokens. A token, here, is a small chunk of language, picture, or action. The largest budget, the post says, corresponds to 100,000 hours of human action recovered from video. As that budget grew, the company reports a power-law drop in next-action error and in whole-body pose error after the model was adapted. A power law, here, means the error keeps falling along a steady curve as the training budget grows, on the company’s own fit. The adaptation sets it names are public egocentric human data — video from a camera worn on a person — Unitree G1 teleoperation data, where a person drives the robot, and the company’s own LightBot data. RuntimeWire adds that Light Origins has not put a total video-hour figure on the production model beyond those measured budgets. File the six budgets, the 100,000 hours, and the power-law line as the post’s. File the missing production total as RuntimeWire’s. This desk did not refit the curve.

The founder line is the wire’s, not a line on the model card. RuntimeWire says Roger Jiang previously worked at OpenAI on the stacks behind InstructGPT, ChatGPT, and GPT-4, that he holds a physics doctorate from the University of Maryland, and that he founded Light Origins in late 2024. File that biography as RuntimeWire’s. This desk did not check a résumé.

Two scores stay company measurements. On a simulated kitchen test the post calls RoboCasa GR-1, 24 tasks, Light Origins says Light-O1 reached a 79.3% macro success rate, with 50 tries per task. Macro means the average of the 24 separate success rates. The post says the company trained the comparison models it names GR00T N1.7 and π0.5, and took a third result from that model’s own paper. That is a Light Origins-run comparison, not an outside leaderboard. The GitHub page also prints a text-to-motion score: 78.0 overall on a bench called HY-Motion-Bench, against 74.7 for HY-Motion-1.0 and 61.4 for Kimodo. File both numbers as Light Origins’. This desk did not rerun them.

Do not read this as robots shipping from a store. Do not read the open 6B preview as a system that, by itself, runs a household robot. The download writes a motion. A physical robot still needs a controller, and the G1 example in the repository is a simulation. Do not turn 100,000 hours into a claim about every video the production model saw. Do not turn 6B into a parameter count this desk summed, and do not merge Qwen3.5-4B with Qwen3.5-4B-Base.

Plain English for the rest of the card: whole-body = legs, torso, arms, and hands moving together. weights = the learned numbers inside the model. parameter = one of those numbers. Apache 2.0 = the open-source license on the preview weights. humanoid = a robot shaped roughly like a person. frame = one pose in a motion clip; 20 frames a second is the card’s rate. behavior foundation model = the lower controller that turns an intended movement into motor commands. teleoperation = a person driving the robot to collect examples. power law = an error curve that keeps falling as the training budget grows, on the company’s fit. This filing is the 21 Sep launch of Light-O1 and the open Light-O1-Preview checkpoint. It is not a robot for sale.

PRIMARY here: the Hugging Face model card for LightOriginsHQ/Light-O1-Preview and the Light Origins technical post “Introducing Light-O1,” both dated 21 Sep 2026 — Tier A PRIMARY, the company’s own record — plus the Light-O1 GitHub repository that ships the inference code. RuntimeWire’s same-day article is the wire for the founder biography and for its summary of the company’s X thread. The 6B size label, the Qwen/Qwen3.5-4B base line on the card, the Apache 2.0 license, the reasoning-then-action stream, and the human_action_138_v1 field list are the model card’s. The two-part answer, the frames-by-138 layout, the playground, the GEAR-SONIC and MuJoCo simulation limit, and the 78.0 bench line are the repository’s. The launch wording, the LightBot and G1 captions, the six budgets, the 100,000 hours, the power-law line, and the 79.3% kitchen score are the technical post’s. The Qwen3.5-4B-Base tree summary, the X-thread motion list, the narrower-than-household-demos line, the missing production video-hour total, and the Roger Jiang biography are RuntimeWire’s. NOT claimed: that production robots are shipping today, that the open preview alone runs a household robot, that 100,000 hours is the production model’s full video diet, that this desk summed the parameters or reran a score, a merged spelling of the Qwen base, a stock tip, or investment advice. Distinct from the already-filed ifr-humanoid-sales-7000, china-humanoid-ipo-slowdown, and faraday-future-eai-919.