Skip to content
TrustList
LO
AI Model

Light-O1-Preview

New· 4

Text-driven humanoid motion model from Light Origins, a research preview that needs a separate controller to drive a robot

About Light-O1-Preview

Light Origins describes itself as an embodied-AI company, with offices in Beijing, Shenzhen and Singapore. Light-O1-Preview is its open-weight research preview of a model that turns a written instruction into whole-body humanoid motion. A technical blog post from 21 September 2026 presented it as the public preview of Light-O1, the company's line of embodied foundation models, and its Hugging Face repository had been created a day before.

From instruction to movement

The model takes a natural-language instruction, drafts a brief reasoning trace on what the body has to do, and only then produces a run of whole-body humanoid actions. Those actions use one shared humanoid representation at 20 frames per second: root motion, pelvis height, yaw rate, two hand-open values and 22 joint rotations. Because the representation is common to all, an output can be retargeted to other robots. The wider Light-O1 approach, Light Origins says, begins with pretraining on human actions recovered from internet video and then adapts to particular robots.

Built on a small base

The model is built on Qwen3.5-4B. The weight metadata on Hugging Face indicates around 5.5 billion parameters in BF16, while press reports describe it as a 6B model. Running it needs Linux, Python 3.11 and an NVIDIA GPU compatible with CUDA 13. A browser playground is available on Hugging Face Spaces, and the weights and code are released under Apache 2.0.

Evaluation

On HY-Motion-Bench, Light Origins' own evaluation gives an overall score of 78.0. Against it stand 61.4 for Kimodo and 74.7 for HY-Motion-1.0, and all of these are the company's figures.

What it is not

It is a research preview and not a robot controller. To drive a physical robot you would need a separate behaviour foundation model or low-level controller, and the Unitree G1 example that ships with it runs only in MuJoCo simulation. RuntimeWire observes that the preview is narrower than the complete Light-O1 system in the company's demonstrations, so it should not be expected to reproduce those on its own.

TrustList summary

Light Origins makes Light-O1-Preview, an open-weight research preview, and the company is an embodied-AI firm with offices in Beijing, Shenzhen and Singapore. The model takes a written instruction, writes a short reasoning trace about what the body must do, then generates a sequence of whole-body humanoid actions.

Main features

  • Text-to-motion: converts a natural-language instruction into humanoid motion.
  • Shared motion format: a 20-frames-per-second stream that covers base movement, pelvis height, turning rate, 22 joint angles and two hand states, so it can be retargeted to other robots.
  • Browser playground: a demonstration hosted on Hugging Face Spaces.
  • Simulation example: a Unitree G1 example that runs in MuJoCo simulation.

Clients

Researchers exploring humanoid motion generated from text. Light Origins calls it a research preview rather than a robot controller.

Key facts

  • Base model: Qwen3.5-4B, with about 5.5 billion parameters according to Hugging Face metadata.
  • Requirements: Linux, Python 3.11 and a CUDA 13-compatible NVIDIA GPU for inference.
  • Licence: Apache 2.0 for the weights and code.
  • Release: announced in a technical blog post dated 21 September 2026, with the Hugging Face repository created on 20 September.
  • Own evaluation: the company reports 78.0 on HY-Motion-Bench, against 74.7 for HY-Motion-1.0, a result it measured itself.

Driving a physical robot needs a separate behaviour model or low-level controller.

Summary written by TrustList from the business's own published information, Nov 2026.

Benchmarks & AI stats

BenchmarkOfficialCommunity avg
No benchmark scores yet. Be the first to add one.

“Official” values are editor-approved and feed the ranking. “Community avg” is the mean of member submissions (shown for transparency; it never affects the ranking until an editor approves a value).

Sign in to add a benchmark score for this model.

Trust Score

4/ 100
Trust Score: New

An earned signal from verification, reviews, awards, transparency and engagement - the vendor can't buy it.

Verification
0/100 · 20%
Reviews
0/100 · 30%
Awards
0/100 · 15%
Transparency
18/100 · 20%
Engagement
0/100 · 15%
Joining soon
Recommendations
Coming soon
Complaints
Coming soon

Updated 9/22/2026

Request a demo or quote from Light-O1-Preview

Protected by reCAPTCHA - Google Privacy Policy and Terms apply.

By sending, you agree we may share your request and contact details with the provider once you confirm your email.

Reviews

Write the first review of Light-O1-Preview

Used it? Your experience helps other buyers decide.

Write a review

Questions & answers

No questions yet. Be the first to ask about Light-O1-Preview.

Light-O1-Preview alternatives

Similar listings buyers compare against this one.

Microsoft AI's real-time speech-to-text model for voice agents, transcribing 60 languages with continuous language detection

Microsoft AI's high-volume, low-latency variant of MAI-Voice-2.1 for speech in 23 languages, aimed at voice agents

Perplexity's open-weight decision model that picks among given options for routing and checks in AI systems

Anthropic's smallest and fastest Claude 5.5 model, available on its own API and three cloud platforms

Aleph Alpha's open-weight English-German mixture-of-experts model for self-hosting, released under Apache 2.0 in October 2026

Cantina Security's first open-weight security model, a post-trained GLM-5.3-Flash for vulnerability research and exploit checks

Google's open multimodal embedding model that maps text, images, audio and video into one vector space, under Apache 2.0

Black Forest Labs' image generation and editing model with bounding-box layouts, multiple reference images and 4K output