On November 12, 2026, the MOSS team at Fudan University, together with the Shanghai Innovation Institute and MOSI, released MOSS-Live, a full-duplex real-time interaction model, along with benchmark comparisons against GPT-Live-1, Doubao SeedDuplex, and other models. MOSS-Live leads on all six full-duplex metrics, including an irrelevant-speech false-response rate of just 2%, compared with 68% for GPT-Live-1 and 75% for Doubao. Its multi-turn task completion rate, however, still trails GPT-Live-1.
A 2% False-Response Rate: Leading on All Six Full-Duplex Metrics
Full duplex is about judgment in real conversation: don't jump in while the user is thinking, don't answer voices that aren't talking to you, and tell a backchannel "mm-hmm" apart from a real interruption. MOSS-Live's pause-interruption rate is 1.44%, against 10.37% for GPT-Live-1 and 14.29% for Doubao; its irrelevant-speech false-response rate is 2%, 66 and 73 percentage points lower than the two rivals. Its irrelevant-interruption rejection rate is 68%, versus 21% and 13%. Backchannel, filler-word, and irrelevant-speech robustness are 98.17%, 82.16%, and 78.90% — also the best among the evaluated models.
Tied First on Tool Calling, Still Behind GPT-Live-1 on Multi-Turn Tasks
Being good at chatting is not the same as getting things done. On tool-calling success, MOSS-Live scores 51%, tied with Doubao for the top spot, ahead of GPT-Live-1's 45%; but its multi-turn task success rate is 57.93%, below GPT-Live-1's 67.68% and above Doubao's 17.38%. Speech understanding shows the same trade-offs: Big-Bench Audio accuracy is 88.40%, below Doubao's 90.50%, and multi-turn instruction following is 3.35 out of 5, below GPT-Live-1's 3.72. Post-interruption answer quality (4.88) and entity tracking, correction, and safety (3.63) are the highest, so it recovers better after a conversation is interrupted.
MOSS-Live also supports semantic interruption, real-time video, screen sharing, and streaming simultaneous interpretation, with the foreground handling continuous interaction while the background runs time-consuming tasks: a user can ask about hotels mid-conversation while the lookup runs in the background, then brings the result back into the chat. Multi-person meetings, in-car use, and customer service benefit first — the noisier the environment and the more interruptions, the costlier false responses and interruptions become. All figures above come from the vendor's published evaluation; the official page does not disclose sample sizes, error margins, or full evaluation configurations, and scores from different benchmarks should not be combined into an overall ranking.
MOSS-Live is now available for online trials and API access on the MOSS platform, the team's latest public release since MOSS-Image went open source. The real-time voice space also saw Tencent Hunyuan's Hy Realtime launch on the same day, with both models chasing the "talk while getting things done" scenario. What to watch next is whether third parties can reproduce these full-duplex results, and how the model performs on latency and stability in real meetings and in cars.
