Alibaba DAMO's RynnUnify beats Cosmos 3 Nano by 9.75 points

Alibaba's DAMO Academy has published RynnUnify, a 16-billion-parameter embodied-AI model that runs a language reasoner and a video generator side by side. "We introduce RynnUnify, a 16B unified embodied model that combines strong embodied understanding with long-horizon visual planning through interleaved video–language reasoning," the authors write.

The shape of the thing, per the project's GitHub, is a Mixture-of-Transformers: an 8B autoregressive reasoner that plans the next step, plus an 8B rectified-flow video generator that imagines the scene it would produce. The reasoner "predicts the next subtask, visual description, and duration," and the generator conditions its video on that output through joint attention.

The headline numbers come from benchmarks the team introduces itself, Rynn-VP and Rynn-GEU. Against Cosmos 3 Nano, RynnUnify "improves the Rynn-VP overall score by 9.75 points." The authors also report that the model "achieves the highest average embodied understanding score among the evaluated unified multimodal models," a comparison scoped to the models the paper picks.

Base weights are already live on the group's Hugging Face hub under Alibaba-DAMO-Academy/RynnUnify-Base.