WeeTodd MiniMax H3 MLX for ComfyUI
Collection
Paged and quantized MLX artifacts for WeeTodd Nodes MiniMax H3 synchronized video-and-audio generation on Apple Silicon. • 3 items • Updated
Acknowledge the included model license
Review the license and eligibility requirements before requesting access.
Access is provided only to users who are eligible under every included license. Do not request access if your location or intended use is excluded.
Log in or Sign Up to review the conditions and access this model content.
MiniMax H3 video VAE MLX Q8
Project: This checkpoint is built for WeeTodd Nodes, an MLX-native MiniMax H3 custom-node suite for ComfyUI on Apple Silicon. Installation, workflow examples, decoder behavior, and current compatibility notes are maintained in that project. Browse the companion artifacts in the WeeTodd MiniMax H3 MLX collection.
This repository contains a directly loadable MLX affine-Q8 MiniMax H3 video VAE. WeeTodd Nodes constructs the quantized decoder modules before reading tensors, so the loader does not rebuild a complete BF16 VAE first.
Compatibility
- WeeTodd Nodes commit
e04db30or later - MLX 0.32.0 or later
- MiniMax H3 synchronized latent decoding
Place video_vae_affine_q8.safetensors in the ComfyUI model directory and select that file as the
Component Loader's video_vae override.
For the measured low-memory graph, use it with the paged q8-extended H3 transformer and the paged Qwen3-VL conditioner. This repository supplies the video decoder only; the H3 audio VAE remains a separate requirement.
Quantization
- Format:
minimax-h3-mlx-video-vae - Quantization: MLX affine Q8, group size 64
- Scope: 144 video-decoder transformer projections
- Convolution and unselected tensors retain their converted source precision
- File size: approximately 2.94 GB
Measured result
Complete parameter residency fell from 5.208 GB to 2.943 GB. A 640 by 384, 124-frame decode reduced peak MLX allocation from 7.660 GB to 5.382 GB. Decode time changed from 22.49 to 23.28 seconds. The Q8 result measured 60.32 dB PSNR and 0.999943 cosine similarity against the BF16 decode.
Provenance
- Base model:
MiniMaxAI/MiniMax-H3 - Base revision reviewed:
bfc8ed0353f5a9733be73e6b2c98ec0948195b86 - Source video-VAE SHA-256:
7c1f131492e7eddacaac9069a61b81bdd39de5cc96561e677c5eab1cdce5e522 - Published artifact SHA-256:
53544746b0e1bc70298574ba549d1c97495c31a446343c8e5a638620b2b5b87f - Conversion implementation:
wee-todd/WeeTodd-Nodes
The safetensors metadata records the source configuration, tensor layout, quantization recipe, source identity, and conversion-format version.
License
The weights are a MiniMax H3 Model Derivative. They remain subject to the MiniMax H3 Community
License Agreement in LICENSE. Review its territorial, redistribution, notice, and acceptable-use
requirements before downloading or using this repository.
WeeTodd Nodes is an independent project and is not affiliated with MiniMax.
Hardware compatibility
Log In to add your hardware
Quantized
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support
Model tree for Vayden/MiniMax-H3-Video-VAE-MLX-Q8
Base model
MiniMaxAI/MiniMax-H3