tieubaoca/PaddleOCR-VL-1.6-GGUF

PaddleOCR-VL-1.6 GGUF

GGUF weights for PaddleOCR-VL-1.6, packaged for use with llama.cpp. The repository contains BF16 and Q8_0 language-model variants plus the BF16 multimodal projector required for image input.

Access to this repository is gated and requests are reviewed manually by the repository owner.

Files

File	Purpose	Size	SHA-256
PaddleOCR-VL-1.6-BF16.gguf	BF16 language model	892.4 MiB	4d8a053ab2048c0f738bd0290506982fe491920b4ef4709609cf2b0294c7712c
PaddleOCR-VL-1.6-Q8_0.gguf	Q8_0 language model	475.2 MiB	024a74326ab652dc56eb34cd855dce5fc3f0b720e26b646a141c942d4fc8cb
PaddleOCR-VL-1.6-mmproj-BF16.gguf	BF16 vision projector; required with either language-model file	840.9 MiB	204d757d7610d9b3faab10d506d69e5b244e32bf765e2bab2d0167e65e0a058a

The exact converter revision and quantization command were not retained with these local artifacts. The filenames, GGUF headers, file sizes, and checksums above describe the published files without making additional conversion-provenance claims.

Usage

Use a recent llama.cpp build with PaddleOCR-VL support.

llama-cli

llama-cli \
  -m PaddleOCR-VL-1.6-Q8_0.gguf \
  --mmproj PaddleOCR-VL-1.6-mmproj-BF16.gguf \
  -p 'OCR:' \
  --image document.png \
  --temp 0

Replace the Q8_0 file with PaddleOCR-VL-1.6-BF16.gguf when higher numerical precision is preferred and sufficient memory is available.

Supported element-level prompts documented by the upstream project include:

• OCR:
• Formula Recognition:
• Table Recognition:
• Chart Recognition:
• Seal Recognition:
• Spotting:

llama-server with PaddleOCR

Start the server:

llama-server \
  -m PaddleOCR-VL-1.6-Q8_0.gguf \
  --mmproj PaddleOCR-VL-1.6-mmproj-BF16.gguf \
  --host 127.0.0.1 \
  --port 8080 \
  --temp 0

Then use PaddleOCR 3.6.0 or newer:

paddleocr doc_parser \
  -i document.png \
  --pipeline_version v1.6 \
  --vl_rec_backend llama-cpp-server \
  --vl_rec_server_url http://127.0.0.1:8080/v1

See the official PaddleOCR-VL documentation for installation, pipeline options, output formats, and image-processing details.

Model description

PaddleOCR-VL-1.6 is a compact vision-language model for multilingual document parsing. The upstream model supports document text, tables, formulas, charts, seals, and text spotting. This repository changes the distribution format and provides a quantized option; it does not represent a new training run.

For architecture, training, benchmarks, intended uses, and the technical report, refer to the upstream model card and the official GGUF release.

Limitations and responsible use

• Quantization can change accuracy and output stability relative to the upstream model. Validate the selected variant on representative documents.
• OCR and structured extraction can contain omissions, hallucinations, incorrect reading order, or malformed tables and formulas.
• Performance varies with language, scan quality, handwriting, layout complexity, resolution, and document domain.
• Do not treat extracted content as authoritative without verification in high-impact workflows.
• Users are responsible for respecting privacy, copyright, data-protection requirements, and the upstream license.

License and attribution

The upstream model is distributed under the Apache License 2.0. This repository follows that license. PaddleOCR-VL was developed by the PaddlePaddle/PaddleOCR team; this repository is an independent GGUF redistribution and is not an official PaddlePaddle release.

Citation

Please use the citation provided by the upstream PaddleOCR-VL-1.6 model card and cite llama.cpp when its runtime is used.