The Duo v1 full-duplex world model, released by inclusionAI on 13 November 2026, is usable today only by researchers with a multi-GPU server. Of the checkpoints on offer, the FP8 distill-quant variant (47.0 GB) is the one to start with, and the 153.5 GB causal base is the reference you keep for comparison and as the quantised model's dependency. The code is Apache-2.0, but the weights ship under a different licence that excludes the EU, the UK and the USA, so check that before building anything on them (GitHub, weights collection).
TL;DR
- Duo v1 is a preview release for researchers: four weight sets on Hugging Face plus offline inference code, published alongside an arXiv paper.
- The paper's streaming setup runs on one server with eight GPUs. Nothing in it supports consumer cards or free Colab.
- Pick distill-quant (47.0 GB) for experiments and keep the causal base (153.5 GB) as the reference. The plain distill (2.8 GB) is small because it holds only the four-step distilled Realization weights.
- Code is Apache-2.0. The weights carry the MiniMax H3 Community License, with a territory carve-out and a 20 million USD revenue threshold (weights licence).
- Per-chunk latency is 330 to 650 ms by component sum, which the authors say is not end-to-end latency.
- Last verified: 15 November 2026.
What is a full-duplex world model?
A full-duplex world model generates a video-and-audio world that keeps perceiving you while it runs. It can speak, interrupt or stay silent while you are still talking, instead of waiting for a turn to end. The paper calls this a Reciprocal World Model: intelligence sits inside the generated world instead of being bolted on as an external control signal (arXiv 2611.15131).
Duo v1 implements the idea in two modules. An Intent module, an autoregressive language model, decides each step whether to emit <speak> plus text or <wait>. A Realization module then generates the coherent audio and video. Both share one timeline, with attention masks that stop future user input leaking into earlier decisions.
Interaction happens in chunks of 17 video frames, about 0.71 seconds at 24 fps, per the same paper.
How do the Duo v1 checkpoints compare?
Four sets of weights are published. Sizes below were read from the Hugging Face repositories on 15 November 2026 (causal base, collection).
| Checkpoint | Size | What it is |
|---|---|---|
| Duo-1.0-causal-base | 153.5 GB, 80 files | Reference causal model |
| Duo-1.0-distill | 2.8 GB, 8 files | Four-step distilled Realization weights |
| Duo-1.0-distill-quant | 47.0 GB, 27 files | FP8 (ModelOpt, per-channel scaling) four-step distilled Realization; depends on the base |
| Duo-1.2-causal-base | 153.5 GB, 80 files | Updated causal base |
The distill-quant model defaults to an Euler sampler with 4 steps and a flow shift of 5.0, according to its model card. Fewer sampling steps is the point: a four-step sampler is far cheaper per chunk than a long diffusion schedule, and chunk time is what a live conversation depends on.
Verdict: for anyone with access to the hardware, start with distill-quant and treat the base as the quality reference. Because distill-quant lists the base as a dependency, you need both on disk. Budget for roughly 200 GB of storage before you begin; that is simple addition of the two published sizes, not a figure from the authors.
Can you run the full-duplex world model on one GPU?
No, not with the reference pipeline. The paper describes streaming inference on one server with eight GPUs: the Intent module runs with tensor parallelism across GPUs 0 to 3 (alongside speech recognition), the Realization diffusion transformer spans GPUs 0 to 7, and a streaming VAE with four-way spatial tiling uses GPUs 4 to 7. VAE decoding of one chunk overlaps with Intent inference for the next (arXiv, Section 6).
Nothing in the paper supports a 12 GB desktop card or a free Colab T4 with 16 GB. If you are weighing consumer hardware for open models generally, our RTX versus Mac verdict on Edge0-145B shows how that fit analysis is usually done, and the Runway Praxis-1 GPU verdict covers open video generation. Duo v1 sits well outside both.
The repository also lists accelerated streaming pipelines on the inference/vllm-omni-duo branch (recommended) and the inference/sglang-omni-duo branch (GitHub). Both target the same class of hardware.
How fast and how good is it?
The authors give a per-chunk figure of 330 to 650 ms, the sum of Intent time, generation time and overhead. They state plainly that this is not end-to-end latency, because the first visible output also needs decoding and may wait for the sampling window to close (arXiv). Treat it as a lower bound on what a user would feel.
For quality, the human evaluation covered 30 cases against 6 criteria, 180 checks in total, of which 168 passed (93.3%, arXiv, Section 7.1). All 90 checks on language adherence, speech intelligibility and visual identity stability passed. Emotion matched facial expression and body language in 96.7% of cases, and lip movement aligned with speech in 90.0% (arXiv, Section 7.1). This is a small, author-run study, so read it as a sanity check rather than a benchmark.
What is the licence catch with Duo v1?
The code and the weights are licensed differently. The GitHub repository is Apache-2.0. The LICENSE file inside the weight repositories is the MiniMax H3 Community License Agreement dated 2 August 2026, and the Hugging Face tag reads "other" (weights licence).
Three terms matter to builders:
- Territory. The applicable territory is worldwide excluding the EU, the UK, the Republic of Korea and the USA.
- Revenue. Commercial products or services earning more than 20 million USD a year need prior written authorisation from MiniMax (weights licence).
- Attribution. Commercial products must show "MiniMax H3" prominently in the user interface.
If you are based in or ship to an excluded territory, do not assume Apache-2.0 on the code covers the weights. Read the file and take legal advice before committing engineering time.
Where does Duo v1 fit among open models?
It is a research artefact, closer to an open reference implementation than a drop-in component. For world-model self-hosting trade-offs in a different family, see Pan-2 versus Wan 2.2. For the audio side of open models, which is far lighter to run, see Sov Transcribe-1 versus Voxtral.
The data pipeline behind Duo is also instructive. The authors used Qwen3-ASR-1.7B with Qwen3-ForcedAligner-0.6B for transcripts, Silero VAD for voice activity, and Qwen2.5-Omni-7B for audio-type review, mixing scripted synthetic video calls with recorded conversations on a shared timeline. Duo v2, promised as "coming soon", is meant to condition generation directly on user audio and video (GitHub).
FAQ
Q: Is Duo v1 really open source?
A: Partly. The inference code is Apache-2.0, but the model weights use the MiniMax H3 Community License, which restricts certain territories and large commercial users.
Q: How many GPUs does Duo v1 need?
A: The paper's reference streaming setup uses one server with eight GPUs. The paper gives no support for a single consumer GPU or free Colab.
Q: Which Duo v1 checkpoint should I download first?
A: Start with Duo-1.0-distill-quant (47.0 GB) for experiments. It depends on Duo-1.0-causal-base (153.5 GB), so plan to download both.
Q: Can I use Duo v1 in a commercial product in the UK or EU?
A: The weight licence excludes the EU, the UK, the USA and South Korea from its applicable territory. Read the LICENSE file and get legal advice before relying on it.
Q: What latency does Duo v1 achieve?
A: The per-chunk component sum is 330 to 650 ms. The authors say this is not end-to-end latency, which also includes decoding and sampling-window waits.
This article was prepared with AI assistance and checked against the primary sources linked above. Corrections log: none to date.
Top comments (0)