Instructions to use Avdpro/EchoMimicV3-MLX with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- MLX
How to use Avdpro/EchoMimicV3-MLX with MLX:
# Download the model from the Hub pip install huggingface_hub[hf_xet] huggingface-cli download --local-dir EchoMimicV3-MLX Avdpro/EchoMimicV3-MLX
- Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- LM Studio
- Atomic Chat
EchoMimicV3 MLX checkpoint
This repository contains the pinned, native-MLX checkpoint layout used by the
AI2Apps ai2apps.model.echomimic-v3-mlx Model Worker on Apple Silicon.
The checkpoint combines:
- EchoMimicV3 Flash transformer from
BadToBest/EchoMimicV3at311e176905a8c4c24b240b530488fe636ce4d249; - Wan-Fun VAE, UMT5 encoder export, and OpenCLIP encoder export from
alibaba-pai/Wan2.1-Fun-V1.1-1.3B-InPatfc913c34361f4ec879e2f9c78b4f11ae50a937d1; - Chinese Wav2Vec2 Base from
TencentGameMate/chinese-wav2vec2-baseatcb511a6498884e41e35686f4b4d6c5188e181773; - UMT5 tokenizer metadata derived from
google/umt5-xxl.
The .safetensors VAE, UMT5, and OpenCLIP files are deterministic, tensor-only
conversions of their pinned PyTorch source checkpoints. See
ai2apps-checkpoint.json for exact file sizes, SHA-256 digests, and source
revisions.
This repository contains model weights only. Use them through a compatible EchoMimicV3 implementation and follow the upstream projects' documentation.
License and attribution
EchoMimicV3, Wan-Fun, UMT5, and the AI2Apps conversion metadata are distributed
under Apache License 2.0. The Chinese Wav2Vec2 Base source repository is marked
MIT by ModelScope. See COMPOSITE-LICENSE.txt for the delivered terms and
NOTICE for exact sources and attribution. Upstream names and trademarks remain
the property of their respective owners.
Quantized