Fix PaddleOCR "No available model hosting platforms detected" and "Unknown argument: use_gpu"

Error 1: model download fails

Exception: No available model hosting platforms detected. Please check your network connection.

PaddleOCR 3.x (tested with 3.7.0) downloads its detection and recognition models the first time you create a PaddleOCR(...) object. It checks a list of model hosts and raises this error when none of them is reachable, which is common behind corporate proxies, in firewalled data centres, inside Docker containers without network access, and in regions where Hugging Face is blocked.

Fix A: pick a model source you can reach

PaddleOCR supports four model hosts. Hugging Face is the default. Select another one with an environment variable:

export PADDLE_PDX_MODEL_SOURCE=modelscope   # or: huggingface, aistudio, bos
python your_script.py

Or set it in Python before importing PaddleOCR:

import os
os.environ["PADDLE_PDX_MODEL_SOURCE"] = "modelscope"

from paddleocr import PaddleOCR
ocr = PaddleOCR(lang="en")

In mainland China, modelscope, aistudio or bos are usually faster than Hugging Face.

Fix B: use a proxy

If your network requires a proxy, make sure Python can see it:

export HTTPS_PROXY=http://proxy.example.com:8080

Fix C: run offline with local models

Download once on a connected machine by running PaddleOCR a single time. The models are saved in ~/.paddlex/official_models/. Copy the folders you need to the offline machine and point PaddleOCR at them:

from paddleocr import PaddleOCR

ocr = PaddleOCR(
    text_detection_model_name="PP-OCRv5_server_det",
    text_detection_model_dir="/models/PP-OCRv5_server_det",
    text_recognition_model_name="en_PP-OCRv5_mobile_rec",
    text_recognition_model_dir="/models/en_PP-OCRv5_mobile_rec",
    use_doc_orientation_classify=False,
    use_doc_unwarping=False,
    use_textline_orientation=False,
)

Use the folder names you actually find in ~/.paddlex/official_models/; they tell you the model names. When model names or folders are given, lang and ocr_version are ignored. Turning off the three optional models means you only need to copy the detection and recognition folders.

Error 2: old 2.x arguments

ValueError: Unknown argument: use_gpu

Your code was written for PaddleOCR 2.x. Version 3 renamed or removed several options:

PaddleOCR 2.x	PaddleOCR 3.x
use_gpu=True	device="gpu:0" (or "cpu")
use_angle_cls=True	use_textline_orientation=True
ocr.ocr(img, cls=True)	ocr.predict(img)
result[0][i][1][0] (text)	res["rec_texts"][i]
show_log=False	Removed

use_angle_cls and ocr.ocr() still work in 3.x with a deprecation warning, but use_gpu fails immediately. We confirmed this with 3.7.0.

# 3.x
from paddleocr import PaddleOCR

ocr = PaddleOCR(lang="en", device="gpu:0", use_textline_orientation=True)
for res in ocr.predict("page.png"):
    print(res["rec_texts"])

Note that device="gpu:0" also requires the GPU build of PaddlePaddle (paddlepaddle-gpu), not the default CPU package.

Full guide

How to OCR a PDF with PaddleOCR in Python: installation, reading results, OCRing PDFs and building a searchable PDF.