Unlimited-OCR-MLX-6bit

Unlimited-OCR MLX 6-bit

MLX 6-bit conversion of baidu/Unlimited-OCR, for Apple Silicon inference through mlx-vlm.

Provenance

• Base model: baidu/Unlimited-OCR (MIT)
• Conversion runtime: mlx-vlm 0.6.10
• Quantization: RTN, 6-bit, group size 64
• Converted on Apple Silicon; no fine-tuning or additional training was performed.

Use

pip install -U mlx-vlm
mlx_vlm.generate \
  --model filipelopesmedbr/Unlimited-OCR-MLX-6bit \
  --image page.png \
  --prompt 'document parsing.' \
  --max-tokens 4096

For structured markdown/layout OCR, use document parsing.. Process documents page-by-page.

Validation and limitation

The converted model passed an end-to-end OCR smoke test through a FastAPI service on a rendered laboratory-report page. It has not been fine-tuned.

mlx-vlm does not currently implement Unlimited-OCR's R-SWA constant-memory mechanism. It therefore uses standard full attention and should be used page-by-page instead of relying on the base model's long multi-page memory claim.

License and attribution

The converted weights remain subject to the MIT license of the base model. Please cite Baidu's Unlimited OCR work when using this model.