Interact-0
Intro
Interact-0 is the full-duplex audio model Interact introduced in an official blog post on 17 November 2026. The post is titled Interact-0 and, in the Architecture section, describes a model with 31 billion parameters and 3 billion active, trained on 9 million hours of single-channel audio and 500,000 hours of channel-separated multi-speaker audio, with an audio encoder, a thinking trunk, and text and audio output heads.
The company presents it as a full-duplex model that can handle user and model audio at the same time, not a half-duplex system that waits for a complete utterance before answering. Audio is split into 80 ms ticks. Training first covered single-channel audio continuation, then added text planning so the model can reason about what it will say before speaking, and finally trained it to predict two simultaneous audio streams.
The weights were initialized from Qwen3-Omni. Interact notes that the teacher’s audio encoder used bidirectional attention and consumed a whole clip at once, while Interact-0 needs streaming input, so the team first trained the encoder to use causal attention on ASR tasks until word error rate fell below that of Qwen3-Omni itself, then trained the trunk.
Advertisement
Explain
The official architecture is specific. The encoder turns audio into features for the trunk; the output heads predict text tokens and audio codebooks, and only one head is used at each position. After text planning is added, positions alternate between text and audio, with each text token appearing slightly before it is spoken, and empty text slots filled with PAD and EPAD tokens in the style of Moshi. To train text planning, the team took 2.5 million hours of higher-quality data, transcribed it, and aligned words to the 80 ms grid with a 100 ms buffer before the corresponding audio.
The training objectives also changed across stages. During pure audio continuation, an anchor loss on text was kept so the model would not lose verbal intelligence: 90% KL divergence against Qwen3-Base and 10% cross-entropy on a diverse text corpus. For text planning, audio positions kept pure cross-entropy, while text positions started with the same KL and CE mix. Because spoken and written language follow different distributions, the model quickly surpassed the teacher at predicting the next text token during speech, so the team lowered and then removed the KL term.
Interact reports internal benchmarks for consistency, world knowledge, and logical reasoning, plus speaker-similarity and continuation-loss curves. The post says gains in the pure audio stage were largely in audio quality, voice consistency, and word and sentence formation rather than intelligence; text planning then raised accuracy on the same internal benchmark. The later part of the post turns to the scarce channel-separated data needed for duplex training. This entry follows that official technical write-up and does not treat unofficial checkpoint or product claims as specifications.
Usage
The first-party write-up is interact.inc/blog/interact-0. The differences to keep in mind versus ordinary spoken-dialogue models are the ones the post itself emphasizes: streaming causal encoding on 80 ms ticks, text planning before speech, and full-duplex prediction of user and model audio.
The public post is a model and training report, not by itself a promise of a stable public API or open weights. How to try the model, if it is offered, should be taken from Interact’s product or docs at the time. Because it is initialized from Qwen3-Omni, it should not be described as the same model as Qwen3-Omni.
When reading the results and audio samples, keep Interact’s own stage split: pure audio continuation improved sound and fluency, and the jump on the intelligence-oriented internal benchmark came after text planning was added.