numind-NuExtract-2.0-8B-MLX

This repository contains an unquantized MLX conversion of numind/NuExtract-2.0-8B for Apple Silicon inference with mlx-vlm.

Provenance

• Upstream model: numind/NuExtract-2.0-8B
• Upstream authors: NuMind
• Conversion type: direct MLX format conversion
• Quantization: none
• Conversion tool: mlx-vlm 0.4.2
• Weight format: safetensors
• Approximate repository size: 16.6 GB

Disclaimer

• This is not a new fine-tune, retrain, or modified checkpoint.
• This repository exists only to make the original model usable through mlx-vlm on Apple Silicon.
• Model behavior, capabilities, limitations, benchmarks, and intended-use guidance come from the upstream model card.
• Credit for the original model, training, evaluation, and documentation belongs to NuMind.
• If you want the canonical model documentation, use the upstream repository: numind/NuExtract-2.0-8B.

License

The upstream model card lists the model license as MIT. This MLX conversion is intended to be shared under the same upstream license and terms.

Use with mlx-vlm

import json

from PIL import Image
from mlx_vlm import generate, load

model, processor = load("mlx-community/numind-NuExtract-2.0-8B-MLX")

template = json.dumps(
    {
        "invoice_number": "verbatim-string",
        "total": "number",
    },
    indent=2,
)

messages = [{"role": "user", "content": [{"type": "image", "image": "embedded"}]}]
prompt = processor.tokenizer.apply_chat_template(
    messages,
    template=template,
    tokenize=False,
    add_generation_prompt=True,
)

image = Image.open("document.png").convert("RGB")
result = generate(
    model,
    processor,
    prompt=prompt,
    image=image,
    max_tokens=2048,
    temperature=0.0,
    verbose=False,
)

print(result.text)

NuExtract works best with temperature at or very close to 0.0 for extraction workloads.