MOSS-Live

MOSS-Live is a full-duplex real-time interaction system from 模思智能 (MOSS). It listens and speaks at the same time, and uses semantic cues to decide when to respond, wait or accept an interruption. It tracks earlier context across topic changes and can resume content that was cut off.

It combines speech, live video, screen sharing and ambient sound. Its listed capabilities include rejecting speech not directed at it (such as car navigation prompts), following multi-speaker conversations, streaming simultaneous interpretation, and proactive interventions based on user-set conditions, for example correcting recitation errors.

The architecture splits a foreground interaction model (perception, speech, turn-taking) from a background model (reasoning, tool calls, long-running tasks), linked by a coordination layer with shared memory and task state. Results from background tasks are brought back into the conversation at a suitable moment. The vendor reports leading results on six full-duplex metrics across FDB v1, v1.5, v2 and v3, a 51% tool-call success rate on FDB v3 (tied highest), and 57.93% multi-turn task success on τ-Voice (second). These are vendor-reported figures. Per creator coverage it is available through an online demo and API; no weights have been announced, and pricing was not provided.