AGIBOT unveils GO-3 Preview, an embodied foundation model it says tops three robot benchmarks
The company claims first place on RoboDojo, RoboTwin and RoboColiseum and promises to open-source the model by the end of 2026; the rankings are its own announcement.
Why it matters
GO-3 Preview is built on AGIBOT's World Reasoning Action Model architecture, which combines task understanding, prediction of future states and action generation. The company pairs it with a post-training system and an agent framework so robots keep learning after deployment, an approach it calls full-lifecycle scaling. Chinese humanoid makers are racing to own the base model as well as the hardware. The test is whether the promised open-source release lands and independent groups reproduce the benchmark results.
Key facts
- Model: GO-3 Preview, built on the WRAM architecture
- Claims first place on RoboDojo, RoboTwin and RoboColiseum
- To be fully open-sourced by the end of 2026
- Paired with Genie Evolver post-training and the Genie Master agent framework
- Announced in Shanghai on 3 November 2026