Alibaba's RynnBrain Model Sets AI Benchmarks
Alibaba's embodied intelligence foundation model, RynnBrain, has reportedly broken 16 embodied AI benchmarks, outperforming models from Google and NVIDIA. The release includes seven open-source models designed to help robots better understand space and predict motion.
- The model was developed by Alibaba's DAMO Academy and is built upon their Qwen3-VL vision-language model, featuring a self-developed architecture called RynnScale that reportedly doubled training speed. - RynnBrain's core technical innovations are its spatiotemporal memory and physical-space reasoning, which allow it to recall past states to inform current actions and better understand the physical environment. - The open-source release includes seven models available on Hugging Face and GitHub, with several versions including 2-billion and 8-billion parameter dense models, and a 30-billion parameter Mixture-of-Experts (MoE) variant. - In addition to the base models, Alibaba released three specialized versions: RynnBrain-Plan for manipulation planning, RynnBrain-Nav for navigation, and RynnBrain-CoP for spatial reasoning. - The 30B MoE model reportedly requires only 3 billion active parameters during inference to achieve its high performance, a significant efficiency that enables faster and smoother robot actions. - While specific benchmark names have not been detailed, Alibaba claims RynnBrain has set new records across 16 open-source embodied AI benchmarks in areas like environmental perception, spatial reasoning, and visual question answering. - To further industry-wide evaluation, DAMO Academy has also released a new benchmark suite called RynnBrain-Bench, which is designed to assess fine-grained spatiotemporal embodied AI tasks. - This initiative is part of a broader push into "Physical AI," where China is strategically competing with the US, and it places Alibaba in direct competition with models like Google's Gemini Robotics-ER 1.5 and NVIDIA's Cosmos-Reason2.
