Instructions to use Vayden/MiniMax-H3-Ref2VA-MLX-q8-extended-paged with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- MLX
How to use Vayden/MiniMax-H3-Ref2VA-MLX-q8-extended-paged with MLX:
# Download the model from the Hub pip install huggingface_hub[hf_xet] huggingface-cli download --local-dir MiniMax-H3-Ref2VA-MLX-q8-extended-paged Vayden/MiniMax-H3-Ref2VA-MLX-q8-extended-paged
- Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- LM Studio
- Atomic Chat
Acknowledge the MiniMax H3 license
Review the included license and eligibility requirements before requesting access.
Access is provided only to users eligible under the MiniMax H3 Community License Agreement. Its territorial and use restrictions apply to these derivative weights. Do not request access if your location or intended use is excluded. By requesting access, you agree to the included license, including Section V and Exhibit A.
Log in or Sign Up to review the conditions and access this model content.
MiniMax H3 Ref2VA MLX Q8 extended paged transformer
This is the genuine Ref2VA transformer from the official MiniMax H3 release, converted for WeeTodd Nodes and WeeTodd Studio on Apple Silicon. Download the complete directory to use the existing paged MLX checkpoint without local conversion.
This repository supplies the transformer component for H3 reference-conditioned synchronized audio/video generation. It requires the shared H3 conditioner (including its vision support), processor/tokenizer, video VAE and audio VAE. Reference-task availability depends on the installed WeeTodd runtime. This publication does not establish inference quality or end-to-end task validation.
Format
- Mixed MLX affine Q8 extended profile, group size 64; exact overrides in
quant_config.json. - Core and AdaLN quantization disabled; remaining tensors retain their converted source precision.
- 50 transformer block pages plus one fixed page;
weetodd-h3-paged-v1. - Tensor payload: 58,260,298,240 bytes; file bytes also include safetensors headers.
- Keep
config.json,quant_config.json,paged_manifest.json,pages/, and all accompanying notices together.
Provenance and verification
Official source: MiniMaxAI/MiniMax-H3, Ref2VA/transformer.
Pinned source revision: 42ed227ee7df40d41602854ae760620d6eb651fe.
All 13 native source shard hashes match the pinned official repository. Every published page was
hashed and checked against its manifest. Tensor names, shapes and dtypes match the local converted
checkpoint. See PROVENANCE.json, MODIFICATIONS.md and the per-file notices.
These are packaging and integrity checks; no generation or full model parity test was performed for publication.
License
The original MiniMax H3 Community License Agreement, including its Acceptable Use Policy, is
included unchanged as LICENSE. Its territorial and use restrictions apply. The agreement excludes
the European Union, the United Kingdom, the Republic of Korea and the United States of America from
its Applicable Territory unless separately authorized by MiniMax. Access acknowledgement does not
expand the license or grant an exception.
This is an independent WeeTodd conversion, not an official MiniMax release. Powered by MiniMax H3.
- Downloads last month
- 4
Quantized
Model tree for Vayden/MiniMax-H3-Ref2VA-MLX-q8-extended-paged
Base model
MiniMaxAI/MiniMax-H3