Duo explores real-time audiovisual conversations, with a major license restriction

inclusionAI has released Duo v1, a research system intended to let an AI listen and respond during a continuous audiovisual conversation. Its November 13 release includes model weights, offline inference code and streaming pipelines. The aim is to move beyond the familiar pattern in which a user speaks, stops, and only then receives a response.

There is a crucial licensing distinction: the project's code is Apache 2.0, but the Duo-1.0-causal-base model is distributed under the MiniMax H3 Community License. That license restricts use, copying and distribution in the United States, European Union, United Kingdom and South Korea unless separately authorized. Check the actual license and your circumstances before downloading or using the weights. Open code does not remove those model restrictions.

Two roles on one timeline

The accompanying research paper describes an Intent component that interprets incoming information and a Realization component that generates the assistant's response. The project synchronizes chunks on a shared timeline to support listening and responding without reducing every exchange to a single turn.

The released v1 architecture and future plans should not be confused: the project describes direct user-audio and video conditioning of Realization as a later direction. The repository offers separate vLLM-Omni and SGLang-Omni streaming pipelines as well as an offline research path. Their software and hardware requirements differ.

What to check before trying it

Real-time multimodal generation depends on more than the size of one checkpoint: the complete pipeline, memory, throughput and latency all matter. The published reference setup uses multiple GPUs, so the existence of a quantized file is not evidence that the complete experience will run on a 12 GB graphics card.

This is a promising research direction, not proof that an assistant will reliably understand interruptions or visual context in everyday life. We have not installed or tested Duo. See the model profile and repository profile, plus the publisher's original materials for the exact artifact and terms.