Duo-1.2-causal-base Released for Real-time Multimodal Models

At a Glance
| Item | Value |
|---|---|
| Repository | inclusionAI/Duo-1.2-causal-base |
| Publisher guide | inclusionAI (Ant Group): models and licenses |
| Published | 2026-11-12 |
| License | minimax-h3-community-license |
| Formats | safetensors |
| Paper | arXiv:2611.15131 |
| Source type | Primary source (the publisher itself) |
Values determined by this site’s code when the information was collected. Dates are JST.
Overview
inclusionAI has released the model weights for “Duo-1.2-causal-base," a full-duplex multimodal world model designed for real-time audio and visual interaction. This release provides the causal-base model weights for Duo 1.2 with improved generation quality, packaging various components tailored for real-time processing. It is a model of high interest for engineers looking to build and test real-time multimodal dialogue systems and world models in local environments or on servers.
According to the paper referenced in the model card, “Duo: A Reciprocal World Model for Full-Duplex Interaction," the model is designed based on the “Reciprocal World Model (RWM)" concept, which treats intelligence and physical reality as equally foundational elements of the world. While traditional world models have primarily focused on simulating physical reality, RWM integrates two capabilities: “Intent," which handles cognition in interaction with the world, and “Realization," which generates coherent audio and visual expressions. Through this, it aims to achieve full-duplex interaction that continuously recognizes audio and video in real time while adapting to user interruptions and spontaneous conversations.
Specifications
The architecture and specifications listed in the model card and metadata are as follows:
- Version: Duo 1.2
- Variant: Causal base
- Package configuration: Complete model bundle
- Realization format: PyTorch Distributed Checkpoint (DCP)
- Default sampler: Euler, 50 steps, flow shift 1.0
Performance
The table information published in the “Model Overview" section of the model card is as follows:
| Field | Value |
|---|---|
| Version | Duo 1.2 |
| Variant | Causal base |
| Package | Complete model bundle |
| Realization format | PyTorch Distributed Checkpoint (DCP) |
| Default sampler | Euler, 50 steps, flow shift 1.0 |
The model card and metadata do not contain quantitative comparison tables or evaluation score data against other models via benchmark tests. Additionally, numerical data is not included in Hugging Face’s evaluation results data area (benchmarks).
According to the qualitative explanations in the model card and statements in the paper, Duo 1.2 is reported to achieve improved generation quality compared to previous versions. Specifically, it highlights functionalities such as real-time continuous recognition of audio, video, and text inputs for full-duplex interaction, adaptation to natural interruptions during conversation, and support for use cases like voice- and video-driven guide assistants and interactive games that do not rely on keyboards or mice.
Strengths and Use Cases
Duo-1.2-causal-base is a multimodal world model that continuously recognizes real-time audio, video, and text inputs, enabling full-duplex interaction with users. Function tags include video-generation, audio-generation, world-model, and duo-world-model, making it well-suited for world model processing that handles video and audio generation seamlessly.
According to the paper and model card descriptions, the main strengths and intended use cases of this model are:
- Full-duplex Dialogue and Interruption Handling: By continuously recognizing audio and video inputs, it allows for the construction of dialogue systems that can respond spontaneously to user utterances and flexibly handle interruptions during conversations.
- Multimodal Spatial and Object Manipulation: It supports visual operations and recognition, such as sharing real-world objects through a camera and modifying or reconstructing virtual scenes. It is suited for developing applications such as guide assistants driven primarily by voice or video without a mouse or keyboard, and interactive audiovisual games.
- Comprehensive Functionality via an Integrated Package: The model package houses the Realization backbone and weights responsible for world representation and consistent response output, alongside the Intent model for cognition and decision-making and alignment adapters. Furthermore, it bundles components such as video and audio codecs, speech recognizers, and forced aligners, covering the range of features necessary for a real-time full-duplex system. Upon runtime integration, configurations can be loaded via
realization/complete.json.
How to Get It
The model files for Duo-1.2-causal-base can be obtained from the Hugging Face repository inclusionAI/Duo-1.2-causal-base. The distribution formats used are safetensors and PyTorch Distributed Checkpoint (DCP).
The command to download the complete model set using the Hugging Face CLI is as follows:
hf download inclusionAI/Duo-1.2-causal-base --local-dir weights
Integration code, various configuration files, and detailed operational steps are available from the official GitHub repository by the developers at inclusionAI/Duo.
Note that this model is not a gated model requiring prior application or individual approval procedures. However, the applied license is specified as minimax-h3-community-license. Please check the license terms before downloading and integrating regarding any commercial use conditions and restriction policies.
Can You Run It Locally?
The publisher distributes this model as safetensors.
License — minimax-h3-community-license: A custom license from the publisher. Check the original terms directly, including whether commercial use is permitted.
Compiled by this site’s code from the published formats and the license field. License summaries are not legal advice — check the publisher’s original terms before relying on them.
Our Own Measurements
Measurements We Did Not Take
We have not confirmed that this site meets the commercial-use terms of this model’s license (minimax-h3-community-license). Because this site carries advertising, we did not run the model (no CPU run, answers, quantization comparison, conversion or generation).
Related Articles
- Duo-1.0: Real-Time Audio-Visual Full-Duplex World Models
- Duo-1.0-distill-quant FP8 Quantized Realization Model: 80GB+ VRAM
- SingVL_S1_4B_GGUF 4B Multimodal Decision-Making Model: 4GB+ VRAM
- ZoomGain-9B-GGUF Vision-Language Model: 8GB+ VRAM
What to Read Next
- Formats this model is available in → Safetensors format guide and models
- Learn about the publisher → inclusionAI (Ant Group): models, licenses and articles

