Body: MLX affine 4-bit (group size 64)
Non-quantized leaves / MTP floats: FP16 (M1/M2 friendly)
Vision: FP16 vision_tower.safetensors (grafted from BF16 source model.visual.*)
Includes mtp.safetensors + mtplx_runtime.json (verified on Apple Silicon)
Approx size: ~21GB on disk

Run with MTPLX

bash

mtplx start --model hawhyhb/Qwen3.6-35B-A3B-Uncensored-Heretic-MTPLX-4bit-FP16
# or local path
mtplx start --model ~/Documents/MTPLX/models/Qwen3.6-35B-A3B-Uncensored-Heretic-MTPLX-4bit-FP16
# API only
mtplx quickstart --model hawhyhb/Qwen3.6-35B-A3B-Uncensored-Heretic-MTPLX-4bit-FP16 --profile sustained --port 8000

Vision works via OpenAI-compatible image_url content parts (no llama.cpp mmproj).

Notes

First forge pass dropped the vision tower; this revision grafts it back and ships preprocessor_config.json / processor_config.json / video_preprocessor_config.json.
Uncensored Heretic text behavior comes from the source repo; use responsibly.