will702/unlimited-ocr-mxfp8-mlx

Unlimited-OCR MXFP8 (mlx-vlm ≥ 0.6 ready)

Drop-in MLX pack of sahilchachra/unlimited-ocr-mxfp8-mlx with configs fixed for native Unlimited-OCR loading on mlx-vlm 0.6+.

What changed vs the upstream MXFP8 pack

File	Upstream (old shim)	This repo
config.json model_type	deepseekocr	unlimited-ocr
processor_config.json processor_class	DeepseekOCRProcessor	UnlimitedOCRHFProcessor
processor_config.json sft_format	deepseek	unlimitedocr
Weights	MXFP8	unchanged

On mlx-vlm 0.6+, the upstream deepseekocr shim often generates repetitive garbage. Routing to unlimited-ocr restores correct OCR.

See FINDINGS.md for details and benchmarks.

Usage

pip install -U "mlx-vlm>=0.6.0" pymupdf

from mlx_vlm import load, generate
from mlx_vlm.prompt_utils import apply_chat_template
from mlx_vlm.utils import load_config

model_id = "will702/unlimited-ocr-mxfp8-mlx"  # update if your namespace differs
model, processor = load(model_id)
config = load_config(model_id)

prompt = apply_chat_template(processor, config, "Free OCR.", num_images=1)
out = generate(
    model, processor,
    prompt=prompt,
    image="page.png",
    max_tokens=4096,
    temperature=0.0,
    repetition_penalty=1.05,
)
text = out.text.replace("Ġ", " ").replace("Ċ", "\n")
print(text)

Prefer prompt Free OCR. — document parsing. often stops immediately on this MLX path.

Credits

• Base model: baidu/Unlimited-OCR (MIT)
• MXFP8 quantization: sahilchachra/unlimited-ocr-mxfp8-mlx
• Runtime: mlx-vlm
• Config fix + Mac tooling: will702/unlimited-ocr-mlx