MOSS-Live ranks first on all six full-duplex voice metrics but trails GPT-Live-1 on τ-Voice multi-turn tasks

MOSI Intelligence and Fudan University's OpenMOSS team report that MOSS-Live, their full-duplex real-time voice system, ranked first on all six reported full-duplex metrics. It also trails OpenAI's GPT-Live-1 on multi-turn task benchmarks, according to the same evaluation. The OpenMOSS blog post, dated November 13, compares MOSS-Live with GPT-Live-1, GPT Realtime 2.1, ByteDance's Doubao SeedDuplex and two Qwen Audio 3.0 models, as reported by Pandaily.

All figures are vendor-published. The post notes that sample sizes, uncertainty estimates and full test settings were not disclosed.

Full-duplex results

MOSS-Live led on every full-duplex metric reported:

• Responding to speech not directed at it: 2%, against 68% for GPT-Live-1 and 75% for Doubao SeedDuplex.
• Cutting in during a user's pause: 1.44%, against 10.37% and 14.29% for the same two systems.
• Ignoring irrelevant interruptions: 68%, against 21% and 13% for the same two systems.
• Holding course through brief acknowledgments such as "mm-hmm": 98.17%.

Where it falls behind

Other results are mixed:

• FDB v3 tool use: MOSS-Live's 51% success rate tied Doubao SeedDuplex for first and beat GPT-Live-1's 45%. Its 2.36-second response latency in tool scenarios trailed GPT Realtime 2.1 at 1.61 seconds.
• τ-Voice multi-turn tasks: MOSS-Live scored 57.93%, behind GPT-Live-1 at 67.68% and well ahead of Doubao at 17.38%.
• Yielding the floor within two seconds: MOSS-Live scored 52.49%, far below the Qwen Audio models at over 93%.
• Big-Bench Audio understanding: MOSS-Live scored 88.40%, below Doubao's 90.50%.
• Multi-turn instruction following: MOSS-Live scored 3.35 out of 5, against 3.72 for GPT-Live-1. It led on answer quality after interruptions at 4.88.

Architecture

MOSS-Live pairs a foreground model with a background model. The foreground model listens, manages turn-taking and speaks. The background model 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 such as a hotel search runs. The system also supports live video, screen sharing and streaming interpretation.

Availability

MOSS-Live is available in an online playground and through a WebSocket realtime API under the model ID moss-live-1.0. Pricing and context window are not yet disclosed. MOSI Intelligence, incubated by the Shanghai Innovation Institute and the OpenMOSS team, previously released the MOSS-Image and MOVA models.

What this means

The results split along two lines. MOSS-Live's strongest numbers measure conversational behavior: not answering speech aimed elsewhere, not interrupting pauses, and not being derailed by backchannel cues. Its weaker numbers come from task-oriented benchmarks (τ-Voice, multi-turn instruction following) and tool-scenario latency, where GPT-Live-1 or GPT Realtime 2.1 lead. The 52.49% rate of yielding the floor within two seconds, against over 93% for the Qwen Audio models, is a further gap in turn-taking.

Because the figures come from the developers and the post gives no sample sizes, uncertainty estimates or full test settings, the margins cannot be assessed independently. Teams comparing realtime voice systems should weigh which behaviors matter for their use case, since no single system led across every category reported.