A lighter version of the part that generates video and audio
inclusionAI (Ant Ling), the AI unit of Chinese financial giant Ant Group, on Nov. 13 released the weights (trained parameters) for "Duo-1.0-distill-quant," a variant of its world model "Duo 1.0," on the unit's organization page on Hugging Face. According to the model card, Duo is a "full-duplex" multimodal world model that exchanges audio and video with the user and can speak and respond at the same time as the user. The new repository contains weights for "Realization," the part of Duo 1.0 that generates video and audio, distilled so that generation takes four steps and quantized to 8-bit floating point (FP8).
According to Hugging Face's API, the repository was created at 5:09 p.m. Nov. 13 Japan time, and it takes up about 47 gigabytes. The release was reported at 6:23:53 p.m. Nov. 13 Japan time by @aitrackerbot, an X account that tracks the appearance of AI models. An announcement on the unit's official X account, @AntLingAGI, had not been confirmed as of the time of writing.
inclusionAI's duo-1.0-distill-quant weights are now public on Hugging Face. • Full-duplex multimodal world model for real-time audio-visual interaction • FP8-quantized four-step distilled Realization weights https://t.co/w3bOe45qqf
Built from a "thinking" part and a "performing" part
According to a technical report the development team posted to arXiv on Nov. 11, Duo consists of two parts. They are "Intent," which handles decisions about interacting with the world, and "Realization," which generates coherent audio and video behavior. Both keep receiving audio, video and text, and the team says the model can keep a conversation going even when the user interrupts. Beyond conversation, the team says users can show it real objects, change virtual scenes and play games without a mouse or keyboard.
According to the model card, the new weights do not run on their own and are used together with the base model "Duo-1.0-causal-base." Intent, the components that encode video and audio, the speech models and the Realization backbone are in the base bundle, and the new weights are laid over it. The base model's default generation process takes 50 steps, while the new version takes four.
The official GitHub repository organizes Duo's weights by version. Version 1.0 now has all three: the base, the distilled version and the distilled and quantized version. For version 1.2, only the base has been released, with the distilled versions listed as "Coming soon." The next version, 2.0, is planned to pass the user's audio and video directly to Realization to support visual interaction.
| Version | Base | Distilled | Distilled and quantized | Features |
|---|---|---|---|---|
| 1.0 | Released | Released | Released (this release) | Basic full-duplex audio and video conversation |
| 1.2 | Released | Planned | Planned | Quality improvements |
| 2.0 | Planned | Planned | Planned | Visual interaction |
No performance figures, an early release for researchers
The Hugging Face model card, the project page and the paper's abstract contain no benchmark results or other performance figures. They also do not show how distillation and quantization changed generation quality or speed, and performance has not been confirmed. ORIENS did not check the full text of the paper for this article.
The GitHub repository describes Duo v1 as "a preview release for researchers." What was released is a pipeline that processes recordings after the fact, and running it requires two CUDA-capable GPUs. The examples on the project page also come from training data and offline inference, and the team says real-time full-duplex results will be added later.
Weights under MiniMax H3 license, U.S. and EU excluded
According to the model card, the new weights were initialized from Chinese company MiniMax's "MiniMax H3" and further trained by the Duo team. The license text defines MiniMax H3 as a video generation model. As a result, the weights are covered by the "MiniMax H3 Community License Agreement."
The license applies only within the "Applicable Territory," which it defines as the world excluding the EU, Britain, South Korea and the U.S. If commercial products or services generate more than $20 million in yearly revenue, prior written authorization from MiniMax is required. Commercial products must prominently display "MiniMax H3" in their user interface. The license also prohibits using the model or its outputs to improve other AI models, and use for military purposes.
On Nov. 11, MiniMax released the weights for "MiniMax-H3-Image," an image generation model in the same H3 series, under a license that likewise excludes the U.S. and the EU. Duo's code on GitHub, meanwhile, was released under the Apache 2.0 license.
Ant Group's AI unit began offering "Ling-3.2-flash," a language model that also reads images and video, on Nov. 5, and in October released the weights for "Ming-3.0," a set of five speech models.
