diff --git a/profile/__init__.py b/profile/__init__.py
new file mode 100644
index 0000000000..e69de29bb2
diff --git a/profile/commons.py b/profile/commons.py
new file mode 100644
index 0000000000..3f7fa69695
--- /dev/null
+++ b/profile/commons.py
@@ -0,0 +1,15 @@
+import os
+from pdf2image import convert_from_bytes
+from PIL.Image import Image
+
+
+def load_pdf_bytes() -> bytes:
+    dir_path = os.path.dirname(__file__)
+    path = os.path.join(dir_path, "resources", "sample_pdf.pdf")
+    with open(path, "rb") as fh:
+        return fh.read()
+
+
+def pdf_to_image(pdf_bytes: bytes, dpi: int) -> Image:
+    images = convert_from_bytes(pdf_bytes, dpi=dpi, grayscale=False)
+    return images[0]
diff --git a/profile/paddle_onnx.py b/profile/paddle_onnx.py
new file mode 100644
index 0000000000..6986368029
--- /dev/null
+++ b/profile/paddle_onnx.py
@@ -0,0 +1,54 @@
+from paddleocr import PaddleOCR
+
+import os
+import numpy as np
+import timeit
+import click
+
+from profile.commons import load_pdf_bytes, pdf_to_image
+import onnxruntime as ort
+
+
+def _init_model(base_dir_path: str) -> PaddleOCR:
+    sess_options = ort.SessionOptions()
+    sess_options.intra_op_num_threads = 8
+    sess_options.execution_mode = ort.ExecutionMode.ORT_SEQUENTIAL
+    sess_options.graph_optimization_level = ort.GraphOptimizationLevel.ORT_ENABLE_ALL
+
+    return PaddleOCR(
+        use_onnx=True,
+        det_model_dir=os.path.join(base_dir_path, "det.onnx"),
+        rec_model_dir=os.path.join(base_dir_path, "rec.onnx"),
+        cls_model_dir=os.path.join(base_dir_path, "cls.onnx"),
+        ocr_version="PP-OCRv4",
+        rec_batch_num=6,
+        onnx_providers=["CPUExecutionProvider"],
+        onnx_sess_options=sess_options,
+    )
+
+
+@click.command()
+@click.option("--paddle-dir-path", default="/home/play/models/paddle2onnx/")
+def main(
+    paddle_dir_path: str,
+):
+    pdf_bytes = load_pdf_bytes()
+    image = pdf_to_image(pdf_bytes, dpi=200)
+
+    model = _init_model(paddle_dir_path)
+
+    times = timeit.Timer(
+        lambda: model.ocr(
+            np.asarray(image),
+            det=True,
+            rec=True,
+            cls=False,
+        )
+    ).repeat(repeat=4, number=1)
+
+    print(f"Mean time: {np.mean(times[1:])}")
+    print(f"Std time: {np.std(times[1:])}")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/profile/profile_paddleocr.py b/profile/profile_paddleocr.py
index d3cba080a8..6ef9c136b0 100644
--- a/profile/profile_paddleocr.py
+++ b/profile/profile_paddleocr.py
@@ -1,13 +1,12 @@
 from paddleocr import PaddleOCR
 
-from PIL.Image import Image
 import os
 import numpy as np
-from pdf2image import convert_from_bytes
 import timeit
 import cProfile
 import click
 from pydantic_settings import BaseSettings, SettingsConfigDict
+from profile.commons import load_pdf_bytes, pdf_to_image
 
 
 class PaddleSettings(BaseSettings):
@@ -27,18 +26,6 @@ class PaddleSettings(BaseSettings):
     model_config = SettingsConfigDict(env_prefix="PADDLE_")
 
 
-def _load_pdf_bytes() -> bytes:
-    dir_path = os.path.dirname(__file__)
-    path = os.path.join(dir_path, "resources", "sample_pdf.pdf")
-    with open(path, "rb") as fh:
-        return fh.read()
-
-
-def pdf_to_image(pdf_bytes: bytes, dpi: int) -> Image:
-    images = convert_from_bytes(pdf_bytes, dpi=dpi, grayscale=False)
-    return images[0]
-
-
 class TimeitDpi:
     def __init__(self, settings: PaddleSettings) -> None:
         self.settings = settings
@@ -47,7 +34,7 @@ def __init__(self, settings: PaddleSettings) -> None:
         self.std_exec_time = [0] * len(self.dpi_values)
 
     def setup(self) -> None:
-        self.pdf_bytes = _load_pdf_bytes()
+        self.pdf_bytes = load_pdf_bytes()
         self.model = PaddleOCR(**self.settings.model_dump())
 
     def timeit(self) -> None:
@@ -75,7 +62,7 @@ def __init__(self, settings: PaddleSettings) -> None:
         self.dpi_values = [50, 100, 150, 200, 250, 300]
 
     def setup(self) -> None:
-        self.pdf_bytes = _load_pdf_bytes()
+        self.pdf_bytes = load_pdf_bytes()
         self.model = PaddleOCR(**self.settings.model_dump())
 
     def profile(self) -> None:
@@ -95,7 +82,7 @@ def profile(self) -> None:
 
 
 def print_recognition_pred():
-    pdf_bytes = _load_pdf_bytes()
+    pdf_bytes = load_pdf_bytes()
     model = PaddleOCR(det=False, rec=True, cls=False)
     image = pdf_to_image(pdf_bytes, dpi=300)
     ocr_result = model.ocr(np.asarray(image), det=False, rec=True, cls=False)
@@ -107,7 +94,7 @@ def print_recognition_pred():
     "--mode",
     default="timeit",
     help="Mode to run the script in.",
-    choices=["timeit", "cprofile", "print_recognition_pred"],
+    type=click.Choice(["timeit", "cprofile", "print_recognition_pred"]),
 )
 def main(mode: str) -> None:
     if mode == "timeit":
diff --git a/profile/scripts/convert_to_onnx.sh b/profile/scripts/convert_to_onnx.sh
new file mode 100755
index 0000000000..d3f8e4656f
--- /dev/null
+++ b/profile/scripts/convert_to_onnx.sh
@@ -0,0 +1,30 @@
+#!/bin/bash
+
+DET_PATH=~/.paddleocr/whl/det/ch/ch_PP-OCRv4_det_infer/
+REC_PATH=~/.paddleocr/whl/rec/ch/ch_PP-OCRv4_rec_infer/
+CLS_PATH=~/.paddleocr/whl/cls/ch_ppocr_mobile_v2.0_cls_infer/
+SAVE_DIR_PATH=~/models/paddle2onnx
+
+paddle2onnx --model_dir $DET_PATH \
+--model_filename inference.pdmodel \
+--params_filename inference.pdiparams \
+--save_file $SAVE_DIR_PATH/det.onnx \
+--opset_version 14 \
+--enable_onnx_checker True \
+--custom_ops '{"paddle_op":"onnx_op"}'
+
+paddle2onnx --model_dir $REC_PATH \
+--model_filename inference.pdmodel \
+--params_filename inference.pdiparams \
+--save_file $SAVE_DIR_PATH/rec.onnx \
+--opset_version 14 \
+--enable_onnx_checker True \
+--custom_ops '{"paddle_op":"onnx_op"}'
+
+paddle2onnx --model_dir $CLS_PATH \
+--model_filename ch_ppocr_mobile_v2.0_cls_infer/inference.pdmodel \
+--params_filename ch_ppocr_mobile_v2.0_cls_infer/inference.pdiparams \
+--save_file @SAVE_DIR_PATH/cls.onnx \
+--opset_version 14 \
+--enable_onnx_checker True \
+--custom_ops '{"paddle_op":"onnx_op"}'
diff --git a/profile/utilities/__init__.py b/profile/utilities/__init__.py
new file mode 100644
index 0000000000..e69de29bb2
diff --git a/profile/utilities/onnx_quantization.py b/profile/utilities/onnx_quantization.py
new file mode 100644
index 0000000000..96a9ec9800
--- /dev/null
+++ b/profile/utilities/onnx_quantization.py
@@ -0,0 +1,38 @@
+from onnxruntime.quantization import (
+    quantize_dynamic,
+    QuantType,
+    quant_pre_process,
+)
+import os
+import click
+
+
+# TODO: produces error: Incomplete symbolic shape inference from
+# onnxruntime.tools.symbolic_shape_infer line 2932.
+def run(dir_path: str, filename: str):
+    base_name = filename.split(".")[0]
+    model_path = os.path.join(dir_path, filename)
+    infer_model_path = os.path.join(dir_path, f"{base_name}_infer.onnx")
+    quant_model_path = os.path.join(dir_path, f"{base_name}_quant.onnx")
+    quant_pre_process(
+        input_model=model_path,
+        output_model_path=infer_model_path,
+    )
+    quantize_dynamic(
+        model_input=infer_model_path,
+        model_output=quant_model_path,
+        weight_type=QuantType.QInt8,
+    )
+
+
+@click.command()
+@click.option("--dir_path", type=str, required=True)
+@click.option("--det_filename", type=str, required=True, default="det.onnx")
+@click.option("--rec_filename", type=str, required=True, default="rec.onnx")
+def main(dir_path: str, det_filename: str, rec_filename: str) -> None:
+    for filename in [det_filename, rec_filename]:
+        run(dir_path=dir_path, filename=filename)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/requirements_dev.txt b/requirements_dev.txt
index 3090387b37..35ff81cdf1 100644
--- a/requirements_dev.txt
+++ b/requirements_dev.txt
@@ -2,3 +2,6 @@ line-profiler==4.1.3
 pdf2image==1.17.0
 pydantic-settings==2.6.0
 click==8.1.7
+paddlepaddle==2.6.0
+paddle2onnx==1.2.11
+onnxruntime==1.20.0