How Interact launched Interact-0 on X
A full-duplex speech model that knows when to speak in group conversations
What Interact launched

Introducing Interact-0, the first in a new generation of always-on models. Interact-0 is a full-duplex speech model designed around human interaction. It is the only model that knows when to speak in groups, and can be active without barging in when it shouldn’t. We share our https://t.co/o1yBwPWO28
Transcript
How LaunchVideo.com rated the launch
LaunchVideo.com rating
About the launch
Interact-0 is a full-duplex speech model that can stay present in a group conversation. It is built for teams and operators who need a voice system in a meeting or a multi-person room, where the hard part is knowing when to speak. The model keeps listening while other people talk, speaks when it is addressed or has something useful to add, and is trained to stay quiet otherwise. Live voice models already handle prompted, one-person chats, but they still take the floor on pauses that belong to someone else. Interact published the model on November 17, 2026 as the first release in its plan for always-on systems that sit in on discussions. The model has 31 billion parameters, with 3 billion active. Interact initialized it from Qwen3-Omni and trained it on 9 million hours of single-channel audio and 500,000 hours of channel-separated multi-speaker audio, in 80 millisecond ticks. A text-planning step lets it decide what to say before it speaks, and each tick carries separate streams for the people in the room and for the model. In the launch demo, several people talk over one another about a trip while a voice called Emma answers only when someone turns to it. The Orchestra checkpoint is trained to stay quiet most of the time and join when called or when it can contribute. No existing benchmark covered group turn-taking, so Interact built Multi-Bench, which checks whether the model can tell when two people are talking to it versus to each other, and Interruption-Bench, which checks whether it keeps speaking through backchannels such as "uh-huh" and "yeah." The Orchestra checkpoint passed 74.2% of Multi-Bench episodes. The Duet checkpoint kept talking through 65% of backchannels. In the same comparison, GPT Live 1 scored 1.5% and 32%, Gemini 3.8 Live scored 4.5% and 1%, and Phonic 1 scored 3.5% and 52%. Interact is a six-person team in San Francisco. A later model is planned to be an order of magnitude larger and to add vision, and the team already runs an always-on instance on its own hardware in the office conference room.
Tags
Questions
Frequently asked questions
The Interact launch post reached 41K views on X, with 596 likes, 108 reposts, 163 comments, 406 bookmarks.
Interact launched on X on November 17, 2026 with the message "A full-duplex speech model that knows when to speak in group conversations".
Interact-0 is a full-duplex speech model that can stay present in a group conversation.
LaunchVideo.com rated the Interact launch 3.40 out of 5: hook 3/5, editing / creativity 4/5, copy 4/5, sentiment of launch 4/5, distribution strategy 2/5.






