You need to agree to share your Hugging Face information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this model content.

ArchVision-OCR-GGUF-vi

High-performance GGUF artifacts for PaddleOCR-VL-1.6 fine-tuned on Vietnamese documents, handwriting, administrative seals, and table recognition (OTSL).

  • Architecture: PaddleOCRVLForConditionalGeneration (ERNIE-4.5 Text Decoder + SigLIP Vision Encoder)
  • Base model: PaddlePaddle/PaddleOCR-VL-1.6
  • Fine-tuning method: LoRA Rank 16 ($\alpha = 32$, 6.05M text-decoder parameters)
  • Checkpoints: Best-performing checkpoint (step 300 / 1.13 epoch) with 0 token truncations, preserving OTSL table structure without repetition loops.

Models:

  • paddleocr-vl-1.6-vietnamese-bf16.gguf (893 MB): Full bfloat16 text decoder.
  • paddleocr-vl-1.6-vietnamese-q8_0.gguf (476 MB): High-fidelity 8-bit quantized text decoder.
  • paddleocr-vl-1.6-vietnamese-mmproj-bf16.gguf (841 MB): Full bfloat16 SigLIP vision encoder + multimodal projector.

Usage with llama.cpp:

Requires llama.cpp build 10142+ or modern release supporting PaddleOCR-VL.

1. Run llama-server:

llama-server \
  -m paddleocr-vl-1.6-vietnamese-q8_0.gguf \
  --mmproj paddleocr-vl-1.6-vietnamese-mmproj-bf16.gguf \
  --alias paddleocr-vl \
  --jinja -ngl 99 --port 8000

2. Run CLI inference:

llama-mtmd-cli \
  -m paddleocr-vl-1.6-vietnamese-q8_0.gguf \
  --mmproj paddleocr-vl-1.6-vietnamese-mmproj-bf16.gguf \
  --jinja \
  --image /path/to/document.png \
  -p "OCR:" \
  --temp 0 -n 1024 -ngl 99
Downloads last month
-
GGUF
Model size
0.5B params
Architecture
paddleocr
Hardware compatibility
Log In to add your hardware

8-bit

16-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for tieubaoca/ArchVision-OCR-GGUF-vi

Quantized
(29)
this model