dgrauet/void-model-mlx-q4

Int4 quantization (group_size 64, transformer Linear weights only) of dgrauet/void-model-mlx, the MLX conversion of netflix/void-model.

Quantized with mlx-forge (mlx-forge convert void-model --quantize --bits 4).

This is the 32 GB configuration: paired with the q8 base model, a full two-pass BigBen run (30 steps, 13 frames, 352×624) peaks at ~23.7 GB — under the 26.8 GB recommended working set of a 32 GB Apple Silicon Mac. Quality vs the bf16 weights: PSNR ≈ 35.5 dB on the same seed.

Usage

These weights can be used with void-model-mlx:

python -m void_mlx.infer \
    --sample sample/BigBen \
    --pass1 weights/q4/void_pass1.safetensors \
    --pass2 weights/q4/void_pass2.safetensors \
    --base-model /path/to/CogVideoX-Fun-V1.5-5b-InP-mlx-q8 \
    --steps 30 --max-frames 13 --height 352 --width 624 \
    --output result.gif

Keep quantize_config.json next to the weights (the loader also infers bits/group_size from the weight shapes if it is missing).

Related Projects

• void-model-mlx (inference): https://github.com/dgrauet/void-model-mlx
• VideoX-Fun-mlx (engine): https://github.com/dgrauet/VideoX-Fun-mlx
• Base model weights (q8): https://huggingface.co/dgrauet/CogVideoX-Fun-V1.5-5b-InP-mlx-q8
• bf16 variant: https://huggingface.co/dgrauet/void-model-mlx
• q8 variant: https://huggingface.co/dgrauet/void-model-mlx-q8

Files

• config.json (365.00 B)
• quantize_config.json (64.00 B)
• split_model.json (1.40 KB)
• void_pass1.safetensors (4.01 GB)
• void_pass2.safetensors (4.01 GB)