python -m mlx_vlm convert \
  --hf-path opendatalab/MinerU2.5-Pro-2604-1.2B \
  --mlx-path MinerU2.5-Pro-2604-1.2B-mlx-8bit \
  -q --q-bits 8 --q-group-size 64

Note: A tie_word_embeddings bug in mlx_vlm's qwen2_vl config parser was patched locally during conversion (nested text_config keys were being discarded, causing a missing lm_head.weight).

Important: Image Size

The layout pass requires images resized to 1036×1036 before inference. Passing images at their original resolution causes the model to produce malformed output (<|box_start|>1.1.1.1... loop). The content pass (per-block crops) works at natural crop dimensions.

Related Models

bf16: carlesonielfa/MinerU2.5-Pro-2604-1.2B-mlx-bf16
8-bit (this model): carlesonielfa/MinerU2.5-Pro-2604-1.2B-mlx-8bit
4-bit: carlesonielfa/MinerU2.5-Pro-2604-1.2B-mlx-4bit
Original: opendatalab/MinerU2.5-Pro-2604-1.2B

License

Inherits the license from the original model.

carlesonielfa/MinerU2.5-Pro-2604-1.2B-mlx-8bit

MinerU2.5-Pro-2604-1.2B MLX 8bit

Conversion

Important: Image Size

Related Models

License

carlesonielfa/
MinerU2.5-Pro-2604-1.2B-mlx-8bit