Interact-0: a 31B full-duplex speech model that knows when to speak in groups
niloofar_mire · x · 2026-11-18
Startup Interact released Interact-0, which it calls the first of a new generation of always-on models—a full-duplex speech model designed around human interaction, able to know when to speak in group conversations and stay active without barging in.
Key details from the report:
- 31B parameters, 3B active; an audio encoder, a thinking trunk, and dual text/audio output heads. Audio is processed in 80ms ticks, each predicting user audio → user text → model text → model audio.
- Trained on 9M hours of single-channel audio plus 500K hours of channel-separated multi-speaker audio.
- Initialized from Qwen3-Omni weights; the bidirectional audio encoder was retrained for streaming causal attention until its ASR word error rate beat Qwen3-Omni itself, then the trunk was trained with a 90% KL-to-Qwen3-Base + 10% CE objective to preserve verbal intelligence.
- Text planning lets the model reason about what it will say before speaking, and it predicts two simultaneous audio streams (user and model) for true full-duplex interaction.
More from Models
- Asked for "exactly 10 amino acids," Claude hallucinates GPT-7 becoming the 267th Pope — LeoTZ03 · 2026-11-18
- Study finds latest Claude models were trained on protein sequences at scale — LeoTZ03 · 2026-11-18
- Haider: No Chinese Model Is on the Pareto Frontier Anymore — Compute, Not Algorithms, Is the Bottleneck — haider1 · 2026-11-18
- SID.ai unveils agentic search model SID-1: 1.9x recall, 24x faster — baseten · 2026-11-18
- Google Unveils Gemini 4 Hydrogen to Power Task Execution in Gemini — GeminiApp · 2026-11-18
- Top token user vouches for Carbon: strong at xhigh for complex coding — olcan · 2026-11-18