Ant's inclusionAI just put Duo 1.0's quantized distilled weights on Hugging Face

It's a full-duplex audio-video interactive world model, listening and watching while it responds, built for real-time interaction

This release is the FP8-quantized, four-step distilled Realization weights, default sampler Euler, 4 steps, flow shift 5.0

The whole repo is 47 GB across 18 safetensors shards

Note this isn't a standalone model, it's an overlay pack, you need the Duo-1.0-causal-base bundle first and layer it on top

License is the MiniMax H3 Community License, paper arXiv:2611.15131 https://huggingface.co/inclusionAI/Duo-1.0-distill-quant