跳到正文
螞蟻 inclusionAI:HuggingFace 新模型·· 9 天前AI 評分45

inclusionAI/Duo-1.0-distill-quant:Duo 1.0 的 FP8 四步蒸餾權重

inclusionAI/Duo-1.0-distill-quant

AI 導讀

inclusionAI/Duo-1.0-distill-quant 收錄 Duo 1.0 的 FP8 量化、四步蒸餾 Realization 權重,供全雙工多模態世界模型使用。這些權重須配合 inclusionAI/Duo-1.0-causal-base;執行環境需支援 ModelOpt FP8 格式或等效轉換路徑。

正文

Duo

Duo-1.0-distill-quant

A Reciprocal World Model for Full-Duplex Interaction

arXiv Homepage Models Code

Model Overview

Duo is a full-duplex multimodal world model for real-time audio-visual interaction. This repository contains the FP8-quantized four-step distilled Realization weights for Duo 1.0.

Field Value
Version Duo 1.0
Variant FP8-quantized four-step distilled Realization
Package Realization overlay
Realization format ModelOpt FP8, per-channel scaling
Base dependency inclusionAI/Duo-1.0-causal-base
Default sampler Euler, 4 steps, flow shift 5.0

Usage

Download the complete Duo 1.0 causal-base bundle, then overlay these Realization weights into the same directory:

hf download inclusionAI/Duo-1.0-causal-base --local-dir weights
hf download inclusionAI/Duo-1.0-distill-quant --local-dir weights

The shared Intent, codecs, speech models, and Realization backbone are supplied by the causal-base bundle. The consuming runtime must support the packaged ModelOpt FP8 format or an equivalent conversion path.

For code, configuration, and runtime-specific integration instructions, see inclusionAI/Duo.

License

These weights were initialized from MiniMax H3 and subsequently trained by the Duo team. See LICENSE and NOTICE for the applicable terms.

來源:螞蟻 inclusionAI:HuggingFace 新模型 · huggingface.co