Why Nuitka Beats PyInstaller: Faster, Smaller Python Executables

1. Experience with PyInstaller and Nuitka

1.1 Requirements

For a recent project I needed to convert Python code into an .exe file, and I discovered two tools that can package Python projects: PyInstaller and Nuitka.

Both tools satisfy the following needs:

Hide source code . PyInstaller encrypts the source with a key, while Nuitka compiles Python to C++ (producing binary .pyd files) to prevent decompilation.

Easy portability . Users can run the program without installing Python or any third‑party packages.

1.2 Impressions

The biggest takeaways after using the two tools are:

PyInstaller experience is poor:

A deep‑learning project packaged with PyInstaller resulted in an exe of nearly 3 GB because it bundles the entire runtime environment.

Packaging is extremely slow, and the resulting executable starts very slowly.

Nuitka is excellent:

The same project produces an exe of only about 7 MB.

Packaging completes in under a minute and the executable starts instantly.

2. Installing and Using Nuitka

2.1 Installation

Install directly via pip: pip install Nuitka Download Visual Studio 2019 (MSVC) or MinGW64 – any C++ compiler will work.

2.2 Usage Process

For projects with many third‑party dependencies (e.g., torch, tensorflow, cv2, numpy, pandas, geopy), it is best to compile only your own code to C++ and leave the large external packages untouched.

Example directory structure of my demo (using PyQt5 for the UI):

├─utils//source folder1
├─src//source folder2
├─logo.ico//demo icon
└─demo.py//main file

Generate an executable with the following command (debug mode):

nuitka --standalone --show-memory --show-progress --nofollow-imports --plugin-enable=qt-plugins --follow-import-to=utils,src --output-dir=out --windows-icon-from-ico=./logo.ico demo.py

Explanation of the most important flags: --standalone: creates a portable package that runs on other machines without requiring Python. --show-memory --show-progress: displays memory usage and progress during compilation. --nofollow-imports: skips compiling imported modules such as keras, numpy, etc. --plugin-enable=qt-plugins: enables the Qt plugin needed for PyQt5 applications. --follow-import-to=utils,src: specifies the two folders containing the code to be compiled to C++. --output-dir=out: sets the output directory. --windows-icon-from-ico=./logo.ico: sets the generated exe's icon. --windows-disable-console: (optional) hides the console window when the exe runs.

After about one minute of compilation you will see a directory structure like:

├─utils//source folder1
├─src//source folder2
├─out//generated exe folder
    ├─demo.build
    └─demo.dist
        └─demo.exe // final executable
        ├─logo.ico // icon
        └─demo.py // main file

Running the exe may initially raise errors such as no module named torch,cv2,tensorflow because those third‑party packages were not compiled. Copy the required package folders (e.g., numpy, cv2) from your Python site‑packages directory into demo.dist. Once the missing modules are added, the exe runs perfectly.