AgiBot releases WITA-Omni embodied omni-modal interaction model

On the evening of November 10, 2026, the Zhiyuan AGIBOT WeChat account announced the official release of WITA-Omni (Silicon Light Dynamic Language full-modal large model). Zhiyuan describes it as the industry's first embodied full-modal end-to-end interactive large model.

According to the company's introduction, WITA-Omni is trained deeply on robot body perception data. In addition to audio and video input and output, it adds new inputs such as tactile sensing and proprioceptive state, and new outputs such as robot motion and facial expressions. Built on a streaming full-duplex architecture, the model can listen, think, respond, act, and make facial expressions simultaneously, cutting the waiting time typical of traditional serial processing. Official figures put the fastest interaction response at about 400 ms and the average at about 600 ms, nearly 3x faster than the industry average, the company claims.

WITA-Omni's full-modal rejection capability removes the need for a wake word: the system combines dialogue orientation, video, audio, voiceprint, lip movement, posture, semantics, and context to decide whether the current utterance requires a response. When multiple people are talking or ambient broadcasts are present, false triggers are greatly reduced. The model supports free switching among hundreds of languages and dialects, and, combined with the full-duplex architecture, enables streaming simultaneous interpretation (including language interpretation and dialect interpretation) as well as bidirectional translation. Its emotional expression is also described as more lifelike, with the robot showing subtle facial changes while listening.

So far, Zhiyuan has not announced when WITA-Omni will be opened up, how it will be accessed, or any commercialization plans, and no independent verification of the performance claims has been reported.