RynnBrain

Alibaba DAMO Academy's open-source embodied 'brain' model, for understanding first-person video, localizing objects, and planning tasks.

RynnBrain was open-sourced by Alibaba's DAMO Academy in February 2026, positioned as a spatiotemporal foundation model that serves as a robot's 'brain' rather than a policy that outputs motor commands directly. It combines four capabilities in one model: understanding first-person video, spatiotemporal grounding (locating objects and operable parts in video), reasoning grounded in physical commonsense, and task planning. Version 1.0 came in 2B, 8B, and 30B-A3B sizes (the last a mixture-of-experts model with 30B total and about 3B active parameters), and was further post-trained into RynnBrain-Nav (navigation), -Plan (planning), -VLA (action), and a spatial-reasoning-focused -CoP variant, alongside the RynnBrain-Bench evaluation suite. Version 1.1, released in July 2026, is built on Qwen3.5 and scales up to 2B, 9B, and 122B-A10B (the largest also a mixture-of-experts model), adding contact-point prediction, with the 2B and 9B versions also supporting native 3D grounding; the company states its largest model beats the closed- and open-source models it evaluated against on three spatial-reasoning benchmarks — VSI-Bench, MMSI, and RefSpatial-Bench. Code and weights are released under Apache 2.0.

ExampleGiven 'put the cup on the table into the sink,' RynnBrain first marks the location of the cup and the sink in the first-person image, then breaks the task into steps — walk to the table, pick up the cup, walk to the sink, put it down — for a lower-level action model to execute.

Also called

RynnBrain 1.0, RynnBrain 1.1, Alibaba DAMO Academy Open Embodied Foundation Models

Related

Alibaba DAMO Academy · Embodied Foundation Model · Embodied Reasoning Model · Brain–Cerebellum Architecture · RynnVLA-002 · Mixture of Experts

Sources

alibaba-damo-academy/RynnBrain (GitHub)
RynnBrain: Open Embodied Foundation Models (arXiv 2602.14979)
RynnBrain 1.1: Towards More Capable and Generalizable Embodied Foundation Model (arXiv 2607.17977)

As of

2026-07