Omni Interaction Agent Technical Report
Gander is presented as a single model for streaming video, speech, and text interaction with agent-style task handling.
The report describes full-duplex use: users can interrupt it, and the model can ask follow-ups or give intermediate feedback. Its architecture splits realtime conversation from higher-level reasoning through a “Cerebellum-Brain” setup. The authors say evaluations cover conversation, omni understanding, interactivity, and agentic intelligence, with internal human tests showing competitive omni interaction and robust behavior in noisy, multi-party, and backchannel-heavy settings. They also say the models, code, and data are being released for research use. HF Daily Papers' note
The report describes full-duplex use: users can interrupt it, and the model can ask follow-ups or give intermediate feedback. Its architecture splits realtime conversation from higher-level reasoning through a “Cerebellum-Brain” setup. The authors say evaluations cover conversation, omni understanding, interactivity, and agentic intelligence, with internal human tests showing competitive omni interaction and robust behavior in noisy, multi-party, and backchannel-heavy settings. They also say the models, code, and data are being released for research use. HF Daily Papers' note
score 5