RynnUnify
Alibaba DAMO's 16B embodied model pairing an 8B reasoner with an 8B video generator for visual planning.
This entry is marked for review. Check the original sources before relying on its details.
Overview
RynnUnify is a 16B unified embodied model from Alibaba's DAMO Academy and Hupan Lab, described in arXiv paper 2610.32965 (October 23, 2026). A Mixture-of-Transformers design combines an 8B autoregressive reasoner that predicts the next subtask, a visual description and a duration with an 8B rectified-flow video generator that renders the predicted scene. An RL-trained agentic extension decides when to consult video predictions (+6.0 points on Rynn-GEU) and a world-action extension targets robot control (51.5% on 20 RoboCasa365 composite tasks, 46.7% across three real-robot tasks). All results are author-reported, including a 9.75-point Rynn-VP gain over Cosmos 3 Nano on benchmarks the same team created.
This is a sourced directory entry. Features and capabilities are reported in the linked sources; this entry does not document independent Aymorphic testing.
Key features
- 16B Mixture-of-Transformers: 8B reasoner plus 8B video generator
- Interleaved video-language reasoning for long-horizon visual planning
- RL-trained agentic extension that chooses when to use video predictions
- World-action extension for robot control
- Author-reported 9.75-point gain over Cosmos 3 Nano on Rynn-VP
Use cases
- Long-horizon task planning for household and industrial robots
- Embodied-agent research on video-based reasoning
- Simulated manipulation benchmarks such as RoboCasa365
Pricing
No hosted access or pricing. Weights availability is unconfirmed: one source cites RynnUnify-Base on Hugging Face, while a release tracker says the paper links no weights, code or API.
Pricing reflects the source record and may change. Confirm current terms with the provider.
Dates and sources
- Released
- 23 Oct 2026
- Record updated
- 2 Nov 2026
- Record last checked
- 2 Nov 2026