Playbooks · Coding
MOSS-Live: full-duplex spoken AI
Written November 10, 2026
MOSS-Live is a spoken AI that listens and talks at the same time, so a conversation can be interrupted, corrected, and resumed without the turn-taking pause. It also keeps working on a task in the background while the conversation goes on.
The problem
Most voice agents take turns: you speak, they answer, and anything you say mid-answer waits or gets lost. A real conversation does not work that way. A short "mm-hmm" means keep going, a pause may mean someone is thinking, and a new question can interrupt an explanation. A spoken assistant that cannot tell those apart either talks over you or waits too long. MOSS-Live, from the OpenMOSS team, is built around that problem: full-duplex speech, live visual understanding, and background task execution in one continuous conversation.
What it does
- Full-duplex conversation. Listen and speak at the same time, handle genuine interruptions, and tell a thinking pause or a brief "mm-hmm" from a request to take the floor.
- Selective attention. Use conversational context to tell speech directed at the assistant from background speech and nearby conversations.
- Continuity across turns. Follow corrections and topic changes, keep the relevant context, and return to an interrupted explanation.
- Live visual understanding. Bring camera or screen-sharing input into the same conversation, alongside speech and ambient sound.
- Tools without a conversational pause. Send research, reasoning, and tool use to the background while the foreground stays available to talk. Ask for hotel information, then ask about parking while that lookup is still running, and the hotel answer arrives when it is ready.
The design splits the work in two: a foreground that runs on the timescale of the conversation, and a background that runs on the timescale of the task. A shared coordination layer connects them through memory, context, and task state.
What the authors measured
The post compares MOSS-Live with GPT Realtime 2.1, GPT-Live-1, Doubao SeedDuplex, and two Qwen Audio 3.0 models across six benchmarks and 36 metrics. The figures below are the authors', from that comparison:
- On FDB v3, MOSS-Live reaches 51.00% tool-use success, tied for the best score in the comparison. On τ-Voice, it reaches 57.93% multi-turn task success, the second-highest score.
- On the full-duplex measures, it has the best score on all six interaction metrics the post charts, including post-interruption response quality at 4.88 out of 5.
- It is not fastest everywhere. Response latency in tool-use scenarios is 2.36 seconds, against 1.61 seconds for GPT Realtime 2.1. On turn-taking latency it is faster: 1.14 seconds against 1.50.
- On Big-Bench Audio overall accuracy it scores 88.40%, behind Doubao SeedDuplex at 90.50% and GPT Realtime 2.1 at 90.10%.
Three worked examples in the post show the shape of it: joining a group meeting and summarising decisions, replanning a road trip while hotel and ferry questions come in, and guiding an English-speaking visitor around Shanghai through the live camera.
Trying it
MOSS-Live can be tried in the online playground linked from the post, or integrated through its realtime API guide. The post also links the code repository. Plan a trip, change your mind halfway through, or turn on the camera and talk about what you see — that is the authors' own suggestion for a first try.
Limits
- Every number on this page comes from the authors' own benchmark comparison. No independent latency test of MOSS-Live exists, and nobody outside the project has confirmed the figures.
- A benchmark table is not a live call. The latency figures are measured on test sets, and a real conversation over a real connection will differ.
- The comparison covers six models the authors chose. A model that is not in the table may beat MOSS-Live on any of these measures.
Related playbooks
- Check whether your session is really on Muse Wonder or Wonder Full — and which route you are paying through
- Cursor builds live apps from cloud agents
- Claude Code 2.1.329 warns where a folder's hooks live
- Addy Osmani's 25 agent skills for coding agents
Sources
Back to: Coding playbooks hub · Also: SafetyPulse: catch safety drift in tuning · Project Eve: Claude plays MMORPGs