MOSS-Live brings full-duplex voice and vision to a realtime API
OpenMOSS says MOSS-Live leads its published full-duplex tests, while a foreground model keeps conversation moving as background agents handle longer tasks.
OpenMOSS has published MOSS-Live, a realtime model and API built around full-duplex conversation: it can keep listening while speaking, handle interruptions, process live audio and vision, and continue longer-running tool work in the background.1
The company’s launch article frames MOSS-Live around a practical problem with voice assistants. Real conversations contain pauses, acknowledgements, changes of plan and speech that is meant for somebody else. A useful realtime system has to decide when to keep talking, when to wait and when to yield the floor. MOSS-Live’s published evaluation compares those behaviours with GPT-Live-1, Doubao SeedDuplex and GPT Realtime 2.1.1
The headline result is conversational timing
In OpenMOSS’s own charts, MOSS-Live leads all six reported full-duplex metrics. It records a 1.44% premature-response rate during pauses, compared with 10.37% for GPT-Live-1 and 14.29% for Doubao SeedDuplex. Its false-response rate to irrelevant speech is 2%, versus 68% and 75% for those two systems.1
The model also rejects 68% of irrelevant interruptions, compared with 21% for GPT-Live-1 and 13% for Doubao. On the tau-Voice measures reported by OpenMOSS, it reaches 98.17% robustness to backchannels, 82.16% robustness to vocal fillers and 78.90% robustness to irrelevant speech.1
Those numbers describe a narrower strength than a universal voice-model victory. They suggest that MOSS-Live is particularly focused on conversational control: distinguishing a meaningful interruption from a cough, acknowledgement or nearby conversation, while avoiding the familiar problem of speaking over a user who has not finished.
Tool use is more mixed
MOSS-Live reaches 51% tool-calling success on the FDB v3 comparison, tying Doubao for the highest result shown and placing above GPT-Live-1 at 45%.1 Its tau-Voice multi-turn task-success score is 57.93%, behind GPT-Live-1 at 67.68% but ahead of Doubao at 17.38%.1
That split matters. A model can feel responsive in a live exchange without being the strongest system for a long chain of instructions. OpenMOSS also reports MOSS-Live at 88.40% on Big-Bench Audio, below Doubao’s 90.50%, while its 3.35/5 multi-turn instruction-following score trails GPT-Live-1 at 3.72/5 and Doubao at 3.44/5.1
The published results therefore support a claim about full-duplex interaction and continuity, not a blanket claim that MOSS-Live wins every audio, reasoning or task benchmark.
A foreground model keeps talking while work continues
The product architecture separates immediate interaction from longer-running work.2 A foreground interaction model handles ongoing perception and conversation. A background model carries out complex reasoning, tool calls and extended tasks, while a coordination layer shares memory, context and task state between them.12
In OpenMOSS’s example, somebody can ask for a hotel search and continue discussing their itinerary while the search runs. The result returns to the conversation when it is ready, rather than forcing the user to sit through a silent wait. The same design is presented for live video, screen sharing and simultaneous interpretation.12
The official API guide identifies the model as `moss-live-1.0` and documents a WebSocket realtime endpoint.3 OpenMOSS also links to a playground where developers can try the system, although the launch material does not establish pricing, latency targets or production availability for every region.13
What the benchmarks do not tell us
These are vendor-published comparisons. The source article does not provide sample sizes, uncertainty estimates, complete evaluation settings, an API snapshot or a universal real-world performance guarantee.1 The model labels also refer to the exact systems named in the charts, not automatically to every product or version in the same families.
That leaves several practical questions open. Developers will need to test latency, interruption behaviour, tool reliability, privacy handling, video costs and failure recovery in their own applications. They will also need to establish whether the foreground/background split behaves consistently when a task fails, when a user changes their mind or when several tools compete for the same context.
Our opinion
MOSS-Live is interesting because it targets one of realtime AI’s least glamorous but most visible problems: knowing when not to speak. The reported 2% false-response rate to irrelevant speech and 1.44% premature-response rate are the numbers most likely to matter in a shared room, a meeting or a car, where a voice assistant that reacts to every sound quickly becomes unusable.
The trade-off is clear in the same release. MOSS-Live does not lead the published multi-turn instruction-following or Big-Bench Audio results, and the benchmark evidence comes from OpenMOSS rather than an independent evaluation. Treat it as a promising specialist design for conversational timing, not a proven all-round replacement for every realtime model. The architecture is the stronger story: keeping a live conversation responsive while a separate process handles the boring work is a sensible direction, provided the tooling and reliability match the demo.