Duo v1 turns the world model into a two-way conversation: 17-frame chunks, it watches while it speaks, and it can be interrupted

World models now count who responds as part of the world itself: Duo v1 receives the user's audio, video and text while generating audiovisual output, and can start speaking, be interrupted, or stay silent while the other side talks. Earlier world models mainly simulated the physical side, with users steering the generated world through limited-degree-of-freedom inputs such as keyboard, mouse, gamepad or text prompts, and nothing inside the world perceived the full user state in real time or decided how to respond. The setting is implemented as chunk-wise full-duplex interaction: each chunk holds 17 video frames, about 0.71 s at 24 fps, and both modules condition only on interaction history before that chunk, so user input arriving during response generation can enter later chunks.

The intent module is no longer just an instruction encoder for a diffusion model; it reasons autoregressively over the ongoing interaction and emits a participation decision. In most audiovisual generative models, a language model serves as the diffusion model's instruction encoder, handling instruction understanding without judging whether to speak at this moment. At every chunk, the intent module emits either a <speak> action followed by text or a <wait> action, conditioned on user input so far and its own earlier language; hidden states rather than decoded text alone carry the decision to the generation module, and action-aware rebalancing during training prevents collapse into perpetual waiting.

Both modules sit on one timeline, and module-specific attention masks enforce causality so the generator cannot see future input. In streaming generation, loose time alignment or masks let future frames leak into the current decision through the visual frontend or cross-chunk attention. User speech, user video, intent tokens and assistant media are synchronized on one timeline, and each chunk receives only the conditions available at its decision boundary; temporal patches in the intent module are completed within the chunk, and noisy media tokens in the realization module can attend to the reference, intent features and clean history but not to their own clean targets or any later chunk.

Dual rollout distillation brings inference-time conditions into training, so the four-step streaming generator learns under conditions close to live interaction. Before distillation, the realization module was trained on intent features of recorded assistant text and on recorded media history, whereas in a live exchange both are produced by the system itself. Within each training window, the frozen intent module autoregressively generates the final assistant turn, and the realization module conditions on the resulting intent features and its own generated audiovisual history; distribution matching distillation on these rollouts yields a four-step streaming generator.

The demonstrated scope is a human evaluation on 30 real-time duplex interaction cases, each reply replayed at normal speed with sound and judged against six binary criteria; the system runs on a single server with eight GPUs, and each chunk covers about 0.71 s of captured audio and video. The report describes Duo v1 as preliminary and currently supporting only basic audiovisual conversation.

Content responsiveness passed only 73.3% (22/30), with misses concentrated in the long tail of conversation that requires reading intent beyond literal words and in cases where intent becomes clear only when speech and video are read together. The report also states that response timing, consistency over longer sessions and recovery from misunderstandings still require broader evaluation, and that an updated version is in development.