mineru-magika

A Python-only fork of Google Magika, based on the published 1.0.3 source. The distribution is named mineru-magika; the Python import remains magika. This fork is maintained independently and is not an official Google release.

Installation

Choose one ONNX Runtime distribution yourself. This package only installs click>=8.1.7; it does not install ONNX Runtime, DirectML, or GPU extras.

# CPU runtime (choose a version compatible with your Python/platform)
pip install mineru-magika onnxruntime

# Windows x64, Python 3.10–3.14
pip install mineru-magika "onnxruntime-directml>=1.23,<2"

Python >=3.8 is supported when a compatible ONNX Runtime is installed. Without a runtime, installation succeeds but importing magika fails with a missing onnxruntime module. There is no runtime download or automatic backend selection.

Do not install magika and mineru-magika together: they own the same Python namespace. In an existing environment, uninstall magika before installing this fork. A downstream requirement for the distribution magika must also be changed; installing this fork does not satisfy that requirement automatically.

Python API and CLI

from magika import Magika

result = Magika().identify_bytes(b"function log(msg) {console.log(msg);}")
print(result.output.label)  # javascript

magika-python-client --version
magika-python-client path/to/file

As in the upstream pure-Python wheel, magika prints a warning directing you to magika-python-client and exits with status 1. No Rust executable is bundled. The Python API, model, configuration and classification behavior are unchanged. The upstream CPUExecutionProvider selection is retained, including when using the DirectML distribution; this fork does not enable GPU inference.

Fork changes and provenance

Version 1.0.3 uses upstream commit 9f225aa480e675af44343b9160f077073ed1b752. Changes are limited to distribution metadata, caller-managed runtime dependencies, distribution version lookups, Python-only packaging, documentation and CI. Model and configuration SHA-256 values are recorded in python/scripts/resource_hashes.json in the source repository and checked against installed artifacts. Original source copyright notices and the Apache-2.0 license are retained. The upstream multi-language sources remain for provenance but are not included in the Python distributions.

Development and release

From the repository's python directory:

uv sync --locked
uv pip install onnxruntime
uv run --no-sync pytest tests
uv build

CI builds a universal wheel and sdist, rebuilds the wheel from the sdist, and tests normal installation without any runtime. CPU checks cover Windows, Linux and macOS on Python 3.10–3.14, plus Linux 3.8/3.9. Windows x64 also checks environments containing only the DirectML distribution. Linux Python 3.12 runs the upstream Python regression suite. Runtime dependencies are installed explicitly by CI.

Tags matching v<version> publish the exact tested artifacts through GitHub Trusted Publishing. Configure the PyPI pending publisher with project mineru-magika, owner myhloli, repository mineru-magika, workflow publish.yml and environment pypi. Publishing requires all validation jobs to pass.