The published model-*.safetensors shards are a server-side copy of the official Mk1 weights. modilify_mlx.load remaps PyTorch parameter names at load time, so you do not need a second converted 49GB checkpoint on the Hub.

Layout

modilify_mlx/ — config, latent deliberation, exclusive commit policy, trunk wrapper, generate loop
convert_modilify_mk1_mlx.py — official safetensors → MLX shards
generate_modilify.py — standalone text trial inference

The heavy decoder / vision trunk is composed from mlx-vlm's DiffusionGemma layers as an internal dependency. Router top-k and the RMS-capped latent merge are Mk1's own forwards.

Convert

Use the oMLX bundled interpreter only as a Python that already has mlx:

bash

export PYTHONHOME=/Applications/oMLX.app/Contents/Resources/Python/cpython-3.11
export PYTHONPATH=/Applications/oMLX.app/Contents/Resources:/Applications/oMLX.app/Contents/Resources/Python/framework-mlx-base/lib/python3.11/site-packages
/Applications/oMLX.app/Contents/Resources/Python/cpython-3.11/bin/python3 \
  ~/Modilify-Mk1-MLX/convert_modilify_mk1_mlx.py \
  --source ~/Modilify-Mk1 \
  --destination ~/Modilify-Mk1-MLX

Generate

Official Mk1 shards and locally remapped MLX shards both load. Parameter names are remapped in modilify_mlx.load.

bash

/Applications/oMLX.app/Contents/Resources/Python/cpython-3.11/bin/python3 \
  ~/Modilify-Mk1-MLX/generate_modilify.py \
  --model ~/Modilify-Mk1-MLX \
  --prompt "Explain why the sky is blue." \
  --max-new-tokens 128

Default is bf16 (--expert-bits 16). Add --profile to print per-phase denoise timings. Full-graph mx.compile of the decoder is blocked by MoE gather/router slices.

modilify/Modilify-Mk1-MLX

Modilify Mk1 MLX

Layout

Convert

Generate

modilify/
Modilify-Mk1-MLX