What if your next AI assistant wasn't just a voice, but an actual face on a video call? Meet Duo v1, a new open-weight audiovisual world model from Ant Group's AI research team. Unlike traditional AI avatars that wait for you to finish speaking, Duo is designed for continuous, full-duplex interaction. It can listen, observe, generate speech and facial expressions, decide when to respond, and handle interruptions. And the character itself is generated by AI. HOW DOES DUO WORK? Most AI world models focus on generating virtual environments you can explore. Duo takes a different approach. Instead of simulating an entire world for a player to navigate, it models a virtual character that can participate in your world. Its architecture combines two systems: • Intent: processes incoming information, reasons about the conversation, and decides whether to speak or wait. • Realization: generates the virtual character's synchronized video, voice, and facial behavior. The two systems operate on a shared timeline. Every 17 video frames, approximately 0.7 seconds at 24 FPS, Intent makes another participation decision. That allows the system to coordinate conversational behavior rather than simply waiting for complete sentences. WHAT MAKES DUO INTERESTING? • Full-duplex interaction: listen and respond during an ongoing conversation. • AI-generated video avatars: synthesize the character's face, expressions, and voice. • Natural interruptions: react when the user interrupts, or participate before a long statement is finished. • Continuous audiovisual context: use incoming video and speech information to guide responses. • Conversational timing: decide when to talk, pause, or keep listening. • Shared visual experiences: research demonstrations explore practical guidance, collaborative planning, and companionship. • Open research: source code, model weights, and implementation details are publicly available. • Streaming inference: experimental real-time pipelines are available through vLLM-Omni and SGLang-Omni. WHY DOES THIS MATTER? I've already been scheduling meetings with AI agents for quite some time. They manage tasks, prepare reports, answer questions, and help coordinate work. But most of these interactions still happen through text or voice. Now imagine scheduling a literal video call with your AI assistant. You open the meeting. Your agent appears on screen. It sees you, listens to your questions, reacts naturally, and participates in the conversation. And eventually, this technology could extend beyond personal assistants. Think about: • Interactive game NPCs. • AI companions. • Virtual tutors and teachers. • Customer service avatars. • Digital characters for interactive storytelling. • Real-time AI presenters. THE CURRENT LIMITATIONS Duo v1 is a research preview, not a ready-to-use consumer video calling application. The main GitHub branch provides an offline demonstration where you record or upload a video and receive an audiovisual response. That reference setup requires Linux and two CUDA GPUs. Separate branches provide experimental streaming implementations, but deployment is significantly more complex than installing a normal chatbot. Also, the 0.7-second figure is the model's decision interval, not a guarantee of 0.7-second end-to-end response latency. The researchers are planning more advanced visual interaction for Duo v2. The source code is Apache 2.0, but the model weights use the MiniMax H3 Community License, which includes geographic and commercial restrictions. THE BIGGER PICTURE We're moving from AI assistants that exchange messages to AI characters that can participate in an ongoing audiovisual experience. And that's a much more interesting direction than simply putting a talking face on top of an LLM. The real challenge is making these characters responsive, reliable, and useful enough that interacting with them feels natural. Maybe the next generation of AI meetings will involve agents in the most literal sense. You won't just invite them to work. You'll invite them to a video call. PROJECT & SOURCES Duo interactive demos: https://duo-v1.github.io/ GitHub: https://github.com/inclusionAI/Duo Model weights: https://huggingface.co/collections/inclusionAI/duo-v1 Research paper: https://arxiv.org/abs/2611.15131 Would you actually want to video call your AI assistant, or is voice interaction already enough? #DuoAI #AIAgents #AIAvatar