inclusionAI's Duo v1 answers you while you are still talking
The team's own human evaluation passed 168 of 180 checks on 30 real-time duplex conversations, with 330-650ms latency across 8 GPUs.

Key takeaways
- inclusionAI released Duo v1 on November 13, 2026, a preview "Reciprocal World Model" that perceives audio and video continuously while responding in real time.
- Duo v1 processes interaction in chunks of 17 video frames, about 0.71 seconds at 24 fps, and ships as four sets of weights on Hugging Face alongside offline inference code under an Apache 2.0 license.
- inclusionAI's own human evaluation passed 168 of 180 checks, or 93.3%, across 30 real-time duplex interaction cases, reporting 330-650ms latency on 8 GPUs.
- The figures are the team's own and have not been independently replicated, and there is no consumer app, hosted demo or price.
inclusionAI released Duo v1 on November 13, a "Reciprocal World Model" that watches and listens continuously while it answers in real time.
The model weights, offline inference code and two accelerated streaming pipelines went up in the company's GitHub repository and on Hugging Face, alongside a technical report on arXiv dated November 11. Duo v1 is a preview release for researchers, and inclusionAI says the next version, Duo v2, will add visual interaction.
Interaction runs in chunks of 17 video frames, about 0.71 seconds at 24 fps, according to the paper. At every chunk an Intent module decides whether the system speaks, emitting either a <speak> action followed by text or a <wait>. A second module, Realization, then generates synchronized video and stereo audio for that chunk.
inclusionAI describes the system as modular. Intent is a multimodal language model that carries each decision to Realization, a causal audiovisual diffusion transformer. Both perceive audio, video and text continuously, which is what lets Duo v1 take a turn mid-stream, respond while the other side is still acting, or stay quiet.
The Realization module was initialized from MiniMax H3 weights and then trained by the Duo team. The complete Duo-1.0-causal-base bundle carries the Intent model, codecs, speech recognizer and aligner, with a sampler default of 50 Euler steps. The distill and distill-quant releases are four-step overlays on top of it, and Duo-1.2-causal-base lists quality improvements. The model cards tag the weights' license as "other"; the code repository is Apache 2.0.
inclusionAI's own human evaluation reports 168 passes out of 180 checks, or 93.3%, across 30 real-time duplex interaction cases, each judged on six binary criteria at normal playback speed. The model passed all 90 checks on language adherence, speech intelligibility and visual identity stability. Lip movement stayed aligned with speech in 90.0% of cases, and content responsiveness, which the paper calls the hardest criterion, reached 73.3% (22 of 30). The team reports latency of 330-650ms from intent inference through latent generation, with about 40ms of overhead, on a deployment that uses 8 GPUs. Those figures are the team's own and have not been independently replicated.
For creators working in virtual production, synthetic presenters and interactive video, Duo v1 is a look at the world-model wave arriving as something you can run and talk to. It is a research release, not a product. There is no consumer app, no hosted demo and no price, and running the offline pipeline needs Linux, Python 3.10 or newer, FFmpeg and at least two CUDA GPUs.
Duo v1 is available now as research weights on Hugging Face and code on GitHub; the repository's code is Apache 2.0, and the weights carry their own license.
Sources
- github.com - release, weights, pipelines, Apache 2.0 code license, hardware requirements, release date
- arxiv.org - paper title, authors, submission date
- arxiv.org - 17-frame/0.71s chunks, Intent and Realization modules, 93.3% human evaluation, 330-650ms latency, 8-GPU deployment
- huggingface.co - causal-base bundle contents, sampler defaults, weights license listed as other
- huggingface.co - four-step distilled Realization weights, model family
- huggingface.co - paper page, publication date, model collection
- aiweekly.co - secondary account of the release date and the 93.3% figure, credited to Hugging Face