Views
No views yet
@misc{cui2025paddleocrvlboostingmultilingualdocument, title={PaddleOCR-VL: Boosting Multilingual Document Parsing via a 0.9B Ultra-Compact Vision-Language Model}, author={Cheng Cui and Ting Sun and Suyin Liang and Tingquan Gao and Zelun Zhang and Jiaxuan Liu and Xueqing Wang and Changda Zhou and Hongen Liu and Manhui Lin and Yue Zhang and Yubo Zhang and Handong Zheng and Jing Zhang and Jun Zhang and Yi Liu and Dianhai Yu and Yanjun Ma}, year={2025}, eprint={2510.14528}, archivePrefix={arXiv}, primaryClass={cs.CV}, url={https://arxiv.org/abs/2510.14528}}paddleocr_vision.float.onnx, paddleocr-vl-text-q41.gguf, and configs/K1 / configs/K3. The vision input is fixed at 784×784, NCHW, with rescale 1/255, mean [0.5,0.5,0.5], and std [0.5,0.5,0.5].git clone --recursive, sets RISCV_ROOT_PATH and SPACEMIT_ORT_DIR, then runs bash build_spacemit.sh glibc.configs/K1, -t 4; K3 uses AI cores 8–15, configs/K3, -t 8. The platform configs set the matching ep_config affinity.1export MODEL_DIR=/path/to/PaddleOCR-VL-0.9B-SpacemiT LLAMA_DIR=/path/to/llama.cpp-installed ORT_DIR=/path/to/spacemit-ort.riscv64.2.0.6
2export LD_LIBRARY_PATH="$LLAMA_DIR/lib:$ORT_DIR/lib:${LD_LIBRARY_PATH:-}"
3"$LLAMA_DIR/bin/llama-server" -m "$MODEL_DIR/paddleocr-vl-text-q41.gguf" --media-backend smt --smt-config-dir "$MODEL_DIR/configs/K1" -t 4 --host 0.0.0.0 --port 8080 --warmupK1 to K3 and -t 4 to -t 8. POST an image data URL and Read the text and describe the document content in the image. to /v1/chat/completions, with thinking disabled. Smoke tests using humanspeech.jpg returned HTTP 200 on K1 (output began The following image is a picture of a building...) and K3 (began The following image is a picture of a building. The text is not clear...). Use a text-heavy document image for OCR accuracy evaluation. The model and upstream dependencies retain their respective Apache-2.0/project licenses.