MOSS-Live
MOSI Intelligence and OpenMOSS released MOSS-Live, a full-duplex real-time voice system that ranks first on six full-duplex metrics against GPT-Live-1 and Doubao SeedDuplex.
MOSS-Live pairs a foreground model that listens, manages turn-taking and speaks with a background model that handles reasoning, tool calls and longer tasks. The two share memory, context and task state through a coordination layer, so the conversation continues while a task runs. It also supports live video, screen sharing and streaming interpretation.
- Date
- Friday 13 November 2026
- Lab
- MOSI.AI
- Kind
- model
- Access
- closed API
Figures
| Measure | Value | Measured by |
|---|---|---|
| Responds to speech not directed at it | 2% lower is better; GPT-Live-1 68%, Doubao SeedDuplex 75% | company |
| Cuts in during user pause | 1.44% GPT-Live-1 10.37%, Doubao SeedDuplex 14.29% | company |
| Correctly ignores irrelevant interruptions | 68% vs 21% and 13% for the two rivals | company |
| Holds course through brief acknowledgments | 98.17% | company |
| τ-Voice multi-turn task score | 57.93% behind GPT-Live-1 at 67.68%, ahead of Doubao at 17.38% | company |
| Big-Bench Audio | 88.40% below Doubao at 90.50% | company |
| FDB v3 tool use success | 51% tied with Doubao SeedDuplex, GPT-Live-1 45% | company |
Figures are vendor-published and sample sizes and uncertainty estimates were not disclosed. Available in an online playground and a WebSocket realtime API as moss-live-1.0; a GitHub repo exists, but weights availability was not confirmed.
Sources
This record was checked against its sources on 14 November 2026. How we check