inclusionAI, the AI research group associated with Ant Group, released Duo v1 on November 13, 2026: an Apache-2.0 research model that takes a user’s video and speech and answers with generated video and audio, rather than text. It is a preview release, and the developers describe it as a first step toward a later version.
What inclusionAI released
The Duo repository on GitHub describes Duo as “a Reciprocal World Model that treats intelligence and physical reality as equally fundamental to a world.” Its news entries are dated November 13, 2026 and cover the paper, streaming pipelines for vLLM-Omni and SGLang-Omni, and the v1 model weights. The repository lists v1.0 checkpoints (causal base, distilled and quantized) and a v1.2 causal base checkpoint; distilled and quantized v1.2 builds are marked as coming soon, and v2.0 is planned but not released. The accompanying report, arXiv:2611.15131, was submitted on November 11, 2026.
The model is not a text-to-video generator in the usual sense. The repository says it supports basic audiovisual conversation, including interruptions, and that its offline demo takes a recorded or uploaded MP4 or WebM clip and plays back Duo’s audiovisual response. At present user input is processed as video plus speech-recognition text; the repository says a planned v2 will condition the response module directly on the user’s audio and video.
How the design works
The report splits the system into two modules. Intent handles the reasoning behind an interaction, including when the model should take part; Realization produces the audio-visual behaviour. According to the abstract, Duo v1 uses autoregressive modeling for interaction reasoning and decisions, aligns chunk-level conditions and tokens on a shared timeline using attention masks that block access to future inputs, and applies dual rollout distillation to reduce distribution shift in both modules. The authors report that the result supports spontaneous conversation, interruptions, sharing real objects, reshaping virtual scenes, guiding assistants and playing audiovisual games without mouse or keyboard input. These are the authors’ own descriptions; the repository gives no benchmark results or parameter counts, and it lists two CUDA GPUs as a hardware requirement for the demo.
What it means for researchers and research offices
Our assessment is that the release matters less as a capability milestone than as an artefact to evaluate. Weights under a permissive licence let a lab run the model locally and inspect its behaviour, which is the practical meaning of an open weights model. Several points follow for teams considering it:
- Treat the claims as unreplicated. The capabilities above come from the developers’ paper and README. Any use in a study should be preceded by a task-specific evaluation, recorded in an AI evaluation card or equivalent.
- Record checkpoint and pipeline versions. With v1.0 and v1.2 checkpoints and separate serving stacks, results are only comparable if the exact checkpoint, serving framework and settings are logged, in line with computational reproducibility practice.
- Read the documentation as a model card. The README and report play the role of a model card, but they do not state training data provenance in the material summarised here, so research offices should not assume it has been cleared for human-subject recordings.
- Plan data handling before use. The model processes video and audio of the person using it. Any study that feeds it recordings of participants needs the usual ethics and data-protection review before upload or local processing.
What is still unknown
The sources do not report independent evaluations, latency figures, training data details or safety testing. Whether Duo’s interaction quality holds up outside the developers’ demonstrations remains to be shown, and the v2 release the authors describe has no date.
Sources
- inclusionAI/Duo repository README, news entries dated November 13, 2026.
- Duo: A Reciprocal World Model for Full-Duplex Interaction, arXiv:2611.15131, submitted November 11, 2026.
- inclusionAI on Hugging Face, listing Duo-1.0 and Duo-1.2 checkpoints.
