AgiBot launches GO-3 Preview and WITA-OMNI, topping three and four embodied AI leaderboards

AgiBot (智元) has released two models, GO-3 Preview and WITA-OMNI, and says they took first place on three and four embodied intelligence leaderboards respectively. The announcements came at the company's Innovation Day, as reported on October 30, 2026 by QbitAI (量子位), a Chinese-language outlet. This article translates and summarizes that Chinese-language report.

The two models

GO-3 Preview is an embodied foundation model aimed at task execution. According to AgiBot, it ranks first on RoboDojo, RoboTwin and RoboColiseum. It uses an architecture called WRAM (World Reasoning Action Model), which combines task understanding, reasoning, future-state prediction and action generation. AgiBot says that after fine-tuning, robots running it can perform part grasping, alignment and insertion, and complete Lego assembly under agent orchestration. The model also supports cross-embodiment transfer, with the company showing it on its X2 biped humanoid opening doors with both hands and retrieving objects from the floor.

For post-training, AgiBot built Genie Evolver, which combines a hundred real robots, physics simulation and world models. A framework called Genie Master handles task planning and exception recovery during deployment.

WITA-OMNI is an end-to-end interaction model. AgiBot says it ranks first on Daily-Omni, TimeLens-Bench, EmbSpatialBench and MMAU. It models speech, vision and text inputs together with speech, facial expression and action outputs in one model. Key details, per AgiBot:

• Average response time of 600 milliseconds, with the fastest around 400 milliseconds.
• Understands more than 100 languages and multiple dialects, and speaks 29 languages and 7 dialects.
• Supports local interaction without a network connection.
• No wake word: it uses sound direction, voiceprint, face and lip movement, then gaze and tone, to decide who is speaking and when to respond.
• A dual-system design: a front-end Interaction Engine for immediate replies and a background agent for reasoning, planning, tool calls and long tasks.
• Intelligent FaceUI, which shows information dynamically on the robot's face screen.

AgiBot plans full domestic deployment by December 2026, an overseas version in Europe in January 2027, and full overseas rollout by March.

Hardware, OS and the Worken metric

Peng Zhihui (稚晖君), AgiBot co-founder, president and CTO, introduced Worken, a unit for effective intelligent work hours. The formula is: Worken = active robots × autonomous working hours × output efficiency ratio.

AgiBot also announced:

• AHA3: a third-generation hardware architecture in which a single SoC runs motion, interaction and task intelligence. The company claims a 50% reduction in physical size and a 90% improvement in response speed.
• G2 Max: a heavy-load robot with 20 kg single-arm payload, 50 kg peak dual-arm payload and seven-degree-of-freedom force-controlled arms. It has hot-swappable dual batteries.
• Lingqu OS 1.0 (灵渠OS): an embodied-native operating system that supports 15 categories and more than 100 peripherals. AgiBot says end-to-end latency jitter is under 4 microseconds at a 1000 Hz task load. It will be fully open-sourced by the end of this year and donated to the OpenAtom Foundation, with production-validated code and more than 5,000 ecosystem software packages.
• AGILE: a motion intelligence framework, plus a subcutaneous joint liquid-cooling system and a new dexterous hand family.

Deployment figures

AgiBot and Longqi (龙旗) run a 3C production line that began scaled production in March 2026 and, per AgiBot, has run stably for more than 7 months. During a 6-day livestream, robots completed more than 64,824 operations and produced 17,625 tablets at a 99.99% success rate, with an average cycle time of 22 seconds per unit, which AgiBot says matches skilled human workers. In retail, 100 Lingxi X2 robots worked as sales guides at Aishida (爱仕达) stores for 3 days. At Hengqin Chimelong Spaceship Park, nearly 500 robots cover more than 100 roles.

What this means

AgiBot is pairing benchmark results with an operational metric. The leaderboard rankings come from the company's own reporting, and Worken is a company-proposed measure, not an industry standard. The stated deployment data (line uptime, success rate, cycle time) gives buyers numbers to test against, while the open-source release of Lingqu OS will let third parties evaluate the software stack on their own hardware.