Zhiyuan unveils WITA-Omni, claiming industry's first embodied multimodal end-to-end LLM

Zhiyuan AGIBOT announced the release of WITA-Omni, a large model described as the industry's first embodied, multimodal, end-to-end interactive model. This system is trained deeply on robot body perception data, supporting not only audio and video I/O but also tactile and body state inputs, alongside robot actions and facial expressions as outputs, effectively giving the robot an interactive brain for the physical world. WITA-Omni utilizes a streaming full-duplex architecture, enabling simultaneous listening, thinking, responding, moving, and emoting, achieving response times of approximately 400 milliseconds, nearly three times faster than the industry average. Furthermore, its multimodal rejection mechanism allows for natural, wake-word-free conversation by assessing context, audio, lip movements, and posture to minimize false triggers. The model boasts advanced human-like emotional expression, capable of displaying dozens of nuanced emotions with varying intensities. Beyond communication, WITA-Omni can generate visualizations on a facial screen to explain complex topics. It also features proactive interaction capabilities, such as greeting VIPs or waving goodbye when prompted by environmental sensing. Previously, in July of this year, the WITA-Omni Preview model topped the DailyOmni Multimodal Understanding Chart with a score of 85.21, surpassing leading models like Qianwen, Gemini, and Doubao.