模思智能 (MOSS) · Full-duplex voice agent

MOSS-Live

MOSS-Live is a real-time conversational AI from 模思智能 (MOSS) that listens and speaks at the same time and can carry out tasks during the conversation. It uses meaning and context to decide when to reply, wait or accept an interruption. It tells back-channel sounds such as "mm" from real interruptions, and it can ignore speech not aimed at it, such as car navigation prompts. It accepts voice, live video, screen sharing and ambient sound, and also supports streaming simultaneous interpretation and proactive interventions on user-set conditions. A foreground model handles perception, speech and turn-taking. A background model handles reasoning, tool calls and long tasks asynchronously. A coordination layer shares memory, context and task state between them. The page reports results on six benchmarks (FDB v1, v1.5, v2, v3, Big-Bench Audio and τ-Voice). It says MOSS-Live ranks first on six full-duplex metrics among the models compared. It reports a 51% tool-call success rate on FDB v3, tied for highest, and 57.93% multi-turn task success on τ-Voice, second among the models compared. The text gives no licence, release date or version.