inclusionAI, the AI research group associated with Ant Group, has released Duo v1, an open model that takes a user's video and speech and responds with generated video and audio rather than text. The release, dated November 13, arrives with an Apache-2.0 license and a paper submitted to arXiv on November 11. It is a preview, and its authors say so, but it points toward a different interaction style for multimodal AI.

What Duo v1 Does

Most conversational models respond in text, even when they accept images or audio. Duo is built around full-duplex interaction. According to the project's README, it takes in audio and video continuously and responds in real time, which allows spontaneous conversation and interruptions. The authors describe it as a reciprocal world model, meaning the system is meant to treat perception and response as one loop rather than as separate input and output steps.

The authors list example uses that include sharing real objects with the system, reshaping virtual scenes, guiding an assistant, and playing audiovisual games without mouse or keyboard input. These are the authors' own claims. They describe what the system is designed to support, and the repository does not present independent evaluations of those tasks.

Architecture and Release Status

The v1 release is more limited than the framing suggests. Per the README, user video reaches the system only through an intent component, as text and video chunks, and not directly into the response generator. The developers say a planned v2 would condition the response directly on the user's audio and video, which is where visual interaction would become more natural.

The offline demo is demanding. It needs Linux with Python 3.10 or newer, FFmpeg on the system path, and by default two CUDA GPUs, one for the intent and codec stage and one for response generation. Separate Python environments are required for the model and speech components. The available checkpoints include v1.0 causal base, distilled and quantized variants, plus a v1.2 causal base model. Several other builds are marked as coming soon, and v2.0 is not yet released.

Those details matter for anyone planning to test it. This is a research release aimed at people with GPU infrastructure, not a consumer app. The Apache-2.0 license makes it easy to study and adapt, but the hardware bar will limit who can run it on their own systems.

What Is Not Yet Proven

The gap between the vision and the evidence is worth noting. The repository includes no benchmark results and no parameter counts, and the coverage found no independent evaluation. Claims about spontaneous conversation and natural interruptions will need testing by outside researchers before they can be compared with text-based systems or with other real-time voice and video models.

The speech-recognition step is another open question. Because v1 passes video through a text-based intent path, the quality of that pipeline will shape how well the system follows what a user is showing it. That is one reason the authors frame v1 as a first step.

Why It Matters

Duo is significant less for what it can do today than for the design direction it signals. If models can process and generate audio and video in a continuous loop, the interface for AI moves away from typing and toward something closer to working alongside a system. That has implications for robotics, interactive software, education and accessibility.

The open release is also notable. An Apache-2.0 model with published checkpoints and a documented pipeline gives researchers a concrete starting point for real-time audiovisual interaction. Anyone building on it should expect rapid change, since the repository already points to distilled, quantized and streaming variants in development.

For now, the sensible reading is cautious. Duo v1 is an interesting research artifact with clear limits, and its real test will come when independent groups measure how well its conversational and interactive claims hold up.

Sources