By No Code MBA · Facts checked
- Provider
- Alibaba
- Announcement date
- Status at last check
- Announced
- Model family
- Rynn
- Capabilities
- Text, Image, Video
What changed
RynnUnify is a 16B unified embodied model from Alibaba DAMO Academy that combines embodied understanding (spatial reasoning, physical prediction, grounding) with long-horizon visual planning through interleaved video-language reasoning. The paper adds an agentic extension trained with reinforcement learning that decides when to consult video predictions, and a world-action extension for robot control. The authors report a Rynn-VP score 9.75 points above Cosmos 3 Nano, a 6.0 point gain on Rynn-GEU from the agentic extension, and 51.5% success on 20 composite RoboCasa365 tasks.
Access and availability
The arXiv paper (PDF, HTML and TeX source) is the only official material. The abstract page does not link to released weights, code or an API.
Official access documentationWhat you can build
For a robotics or simulation prototype, compare direct reasoning with generated visual evidence on a multi-step task, then check whether the predicted scene changes improve the next action.
Limitations
The paper links no model weights, code or hosted access, so builders cannot use it yet. All results are author-reported, including on two new benchmarks (Rynn-VP and Rynn-GEU) created by the same team, and real-robot success averaged 46.7% across three tasks.