Ant Group Releases Duo v1, a Full-Duplex World Model Built for Real-Time Audiovisual Dialogue
InclusionAI open-sources four model checkpoints covering intent inference and causal diffusion generation, running on a single eight-GPU server

InclusionAI, the open-source AI research lab backed by Ant Group, released Duo v1 on November 13, 2026 — a full-duplex world model that perceives a user's live video and speech while simultaneously generating a responding audiovisual character. Unlike pipeline-based digital human systems, Duo v1 handles both listening and speaking at the same time, allowing either party to interrupt mid-sentence without breaking the interaction.
The release includes an arXiv technical paper submitted November 11, 2026 and four model checkpoints on HuggingFace: a causal base and distilled variant at version 1.0, a quantized version of the distilled model, and a causal base checkpoint at version 1.2. All four are released under the MiniMax H3 Community License, which permits non-commercial use; commercial deployments require written authorization from MiniMax, whose weights formed the initialization point for the Realization Module.
A New Interaction Paradigm
The paper introduces the Reciprocal World Model (RWM) framework, which treats a live conversation as a shared timeline perceived and acted on by both the user and the AI character simultaneously. At every 708-millisecond window — covering 17 video frames at 24 frames per second — the model decides whether to speak or wait and then generates the corresponding audio and video before the window closes.
Duo v1 divides the work between two modules. The Intent Module is a fine-tuned multimodal language model that reads the incoming video and speech-recognition text and emits either a <speak> action with accompanying text or a <wait> action. Its hidden states are passed downstream. The Realization Module is a causal audiovisual diffusion Transformer initialized from MiniMax weights and trained with flow matching; it consumes the Intent Module's hidden states and generates joint video and stereo audio latents for the current chunk.
The two modules run in parallel on a single server with eight GPUs. The Intent Module uses four-way tensor parallelism on GPUs 0–3. The Realization Module's diffusion Transformer runs across all eight GPUs using either tensor or sequence parallelism, while a streaming variational autoencoder (VAE) handles spatial tiling on GPUs 4–7.
Architecture and Latency
Duo v1 represents the user's audiovisual stream through a pair of learned codecs. For audio, a VAE produces 40 latent vectors per second per stereo channel, each with 32 dimensions. For video, groups of 17 frames are encoded into 5 latent frames with 24 dimensions each, with 16×16 spatial reduction; a 2×2 merging step yields 96-dimensional inputs to the Transformer.
End-to-end latency per 708-millisecond chunk — measured on 768×576 video — breaks down as: Intent Module inference at 90–160 milliseconds, joint latent generation at 200–450 milliseconds, streaming VAE decode of a four-frame group at approximately 30 milliseconds, and additional overhead of roughly 40 milliseconds. The total rounds to 330–650 milliseconds per chunk, inside the 708-millisecond window.
The paper describes three training stages. Stage 1 fine-tunes the Intent Module on annotated conversational data with action-aware rebalancing and perturbed histories to handle interruptions. Stage 2 trains the Realization Module conditioned on recorded media while the Intent Module is frozen. Stage 3, called Dual Rollout Distillation, uses live rollouts from both modules to collapse a 50-step diffusion process down to four steps, closing the gap between training distribution and live inference conditions and reducing per-chunk generation time substantially.
Evaluation
The team evaluated Duo v1 on 30 real-time full-duplex cases, checking 180 binary criteria. The overall pass rate was 93.3 percent (168 of 180). Among the six criteria, language adherence, speech intelligibility, and visual identity stability each achieved 100 percent. Audio-visual emotional consistency reached 96.7 percent and lip synchronization scored 90 percent. Content responsiveness — the model's ability to address what the user actually said — was the weakest result at 73.3 percent, a gap the authors identify as the primary area for improvement.
The team notes that all benchmark numbers are self-reported and that the evaluation sample of 30 cases is small. No third-party replication has been posted.
Comparison to Prior Work
Kyutai's Moshi, released in September 2024, demonstrated that full-duplex operation was achievable in a speech-only model, with practical latency of around 200 milliseconds. Duo v1 extends the concept to include synchronized video output — a generated face and voice together — turning the interaction into a live audiovisual conversation rather than audio alone.
The paper also distinguishes Duo v1 from pipeline-based digital human systems such as HeyGen, Synthesia, and D-ID, which assemble separate speech and video models in sequence. The sequential pipeline introduces latency between the user finishing a turn and the character beginning to respond, and prevents true simultaneous overlap. Duo v1's chunk-based timeline eliminates that gap: the model can begin generating a response before the user finishes speaking.
InclusionAI describes Duo v1 as a first step, with the current release processing user input as video plus speech-recognition text. A planned version 2 would condition responses directly on the raw audio and video stream, removing the intermediate transcription step. No release date for version 2 has been announced.
Availability
The four v1.0 and v1.2 checkpoints are available on HuggingFace and the technical paper is at arXiv:2611.15131. A project page is live at duo-v1.github.io and the GitHub repository contains inference and deployment instructions.



