Instructions to use gabrielrocco/MiniMax-H3-Ref2VA-MLX-Serve-8bit with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- MLX
How to use gabrielrocco/MiniMax-H3-Ref2VA-MLX-Serve-8bit with MLX:
# Download the model from the Hub pip install huggingface_hub[hf_xet] huggingface-cli download --local-dir MiniMax-H3-Ref2VA-MLX-Serve-8bit gabrielrocco/MiniMax-H3-Ref2VA-MLX-Serve-8bit
- Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- LM Studio
Modifications to MiniMax H3
The REF2VA transformer and text encoder were converted from BF16 to MLX affine 8-bit weights with group size 64. Gathered tables, precision-sensitive projections, output heads and both VAEs remain dense. No weights were retrained or distilled. See LICENSE for the governing terms.