An eight-GPU compute budget
The streaming numbers are the most concrete part of the paper, and the least portable. Keeping up with a 708-millisecond chunk clock at 768×576 took eight GPUs: all eight perform the Realization module's denoising, and between generation phases GPUs 0–3 run Intent inference for the next chunk while GPUs 4–7 decode the current one. The reported 330 to 650 milliseconds is the sum of Intent inference, latent generation and overhead inside that layout — a per-cycle compute budget. The delay to the first visible output additionally includes the initial decoding time and any remaining wait for the sampling window to close. The GPU model is not stated.
None of this makes the figure dishonest. It makes it a lab figure. The configuration the shipped README documents for offline use is two GPUs, one for Intent and the codecs and one for Realization, across two Python environments, and that path publishes no latency number at all. A reader pricing a duplex avatar demo this week should budget from the two-GPU setup and treat the sub-650-millisecond cycles as what the authors achieved on hardware most teams do not have on a desk.
The authors' own evaluation
The 93.3% is an author-reported result, and it needs its denominator kept attached. Thirty cases, six binary criteria with written failure conditions and no partial credit. Language adherence, speech intelligibility and visual identity stability passed all ninety of their combined checks. Content responsiveness was the weakest criterion at 22 of 30, on cases involving jokes, oblique remarks and intent that only resolves when speech and video are read together; lip alignment reached 90.0%.
A binary rubric with written failure conditions is a checkable instrument, and the failure analysis is specific. But it is the paper's own evaluation: there is no third-party run, and the paper offers no head-to-head numbers against the systems it discusses, including Moshi, Qwen-Omni and Wan-Streamer. Duo v1 has been measured against a rubric. It has not been measured against another model.
The license decides who can use it
The weights are public, ungated and complete. They are not open in the way the code is. The Duo repository includes the MiniMax H3 Community License — ALFW read the LICENSE file in the repository first-hand — and its definitions limit the grant to an Applicable Territory that is worldwide excluding the European Union, the United Kingdom, South Korea and the United States. Anyone in those territories is invited to contact MiniMax about a separate license. That is a route, not a permission.
Read plainly, Canada, Singapore, Japan and Switzerland are within the license's stated territory; other conditions in the license still apply. A team in San Francisco, London, Berlin or Seoul is outside the territorial grant without separate authorization, and should learn that before the 153 GB download finishes rather than after a two-week integration. ALFW read the same territorial structure in the H3 license itself in October; Duo shows the clause operating on a derivative release, which makes it a practical fact about the H3 ecosystem rather than a footnote in one model's paperwork.
Two more caveats belong in any plan. The release describes itself as a preview supporting basic audiovisual conversation. And a v2 is promised to condition the Realization module directly on user audio and video. The fuller interactive version the research describes is the next version, not this one.
Who should care this week
Duplex-interaction researchers get a substantial artifact: a paper with a latency budget, a chunk clock and weights that actually download. Avatar and companion builders get a design parameter worth borrowing without borrowing the model: the 708-millisecond chunk clock that structures when the model can react. Everyone else, and every team in the four excluded territories, is reading a preview. The sensible move is to read the paper now and wait for an independent run, a stated GPU model for the latency figures, or a license conversation, whichever your situation needs first.
Sources
- [1] inclusionAI, Duo-1.0-causal-base model repository and README, Hugging Face (created November 12, 2026; release lines dated November 13, 2026)Read source
- [2] MiniMax H3 Community License Agreement, LICENSE file in the Duo-1.0-causal-base repository (read first-hand)Read source
- [3] Cao et al., "Duo: A Reciprocal World Model for Full-Duplex Interaction," arXiv 2611.15131 (submitted November 11, 2026)Read source
- [4] inclusionAI, Duo code repository (Apache-2.0 code; offline and streaming inference paths; preview and v2 statements)Read source
- [5] OrcaRouter, "Duo-1.0-causal-base: 153 GB of Weights, No Launch Post" (November 13, 2026; repository file inventory and config read)Read source
- [6] AI Weekly, "inclusionAI Releases Duo v1, a Full-Duplex Audiovisual World Model With 93.3% Pass Rate at Sub-650ms Latency" (November 13, 2026)Read source