Fast Base64 implementation

This project is a wrapper on libbase64.

It aims to provide a fast base64 implementation for base64 encoding/decoding.

Installation

pip install pybase64

Usage

pybase64 uses the same API as Python base64 "modern interface" (introduced in Python 2.4) for an easy integration.

To get the fastest decoding, it is recommended to use the pybase64.b64decode with validate=True (or ignorechars=b"") and padded=True when possible.

import pybase64

print(pybase64.b64encode(b'>>>foo???', altchars='_:'))
# b'Pj4_Zm9vPz8:'
print(pybase64.b64decode(b'Pj4_Zm9vPz8:', altchars='_:', validate=True))
# b'>>>foo???'

# Standard encoding helpers
print(pybase64.standard_b64encode(b'>>>foo???'))
# b'Pj4+Zm9vPz8/'
print(pybase64.standard_b64decode(b'Pj4+Zm9vPz8/'))
# b'>>>foo???'

# URL safe encoding helpers
print(pybase64.urlsafe_b64encode(b'>>>foo???'))
# b'Pj4-Zm9vPz8_'
print(pybase64.urlsafe_b64decode(b'Pj4-Zm9vPz8_'))
# b'>>>foo???'

A command-line tool is also provided. It has encode, decode and benchmark subcommands.

usage: pybase64 [-h] [-V] {benchmark,encode,decode} ...

pybase64 command-line tool.

positional arguments:
  {benchmark,encode,decode}
                        tool help
    benchmark           -h for usage
    encode              -h for usage
    decode              -h for usage

optional arguments:
  -h, --help            show this help message and exit
  -V, --version         show program's version number and exit

Full documentation on Read the Docs.

Benchmark

Running Python 3.15.0a8, Apple clang version 21.0.0 (clang-2100.0.123.102), macOS 26.3.1, Apple M1 Max

pybase64 1.5.0 (C extension active - NEON)
bench: altchars=None, validate=True, padded=True
pybase64.encodebytes:     5129 MB/s (135,696 bytes -> 183,309 bytes)
pybase64.b64encode:      16339 MB/s (135,696 bytes -> 180,928 bytes)
pybase64.b64decode:       8821 MB/s (180,928 bytes -> 135,696 bytes)
base64.encodebytes:       2339 MB/s (135,696 bytes -> 183,309 bytes)
base64.b64encode:         2642 MB/s (135,696 bytes -> 180,928 bytes)
base64.b64decode:         2613 MB/s (180,928 bytes -> 135,696 bytes)
bench: altchars=None, validate=True, padded=False
pybase64.b64encode:      16369 MB/s (135,696 bytes -> 180,928 bytes)
pybase64.b64decode:       2480 MB/s (180,928 bytes -> 135,696 bytes)
base64.b64encode:         2573 MB/s (135,696 bytes -> 180,928 bytes)
base64.b64decode:         2583 MB/s (180,928 bytes -> 135,696 bytes)
bench: altchars=None, ignorechars=b'', padded=False
pybase64.b64decode:       2478 MB/s (180,928 bytes -> 135,696 bytes)
base64.b64decode:         2610 MB/s (180,928 bytes -> 135,696 bytes)
bench: altchars=None, ignorechars=b'\n', padded=True
pybase64.b64decode:       2338 MB/s (183,308 bytes -> 135,696 bytes)
base64.b64decode:         2083 MB/s (183,308 bytes -> 135,696 bytes)
bench: altchars=None, ignorechars=b'\n', padded=False
pybase64.b64decode:       2337 MB/s (183,308 bytes -> 135,696 bytes)
base64.b64decode:         2013 MB/s (183,308 bytes -> 135,696 bytes)
bench: altchars=None, validate=False, padded=True
pybase64.b64decode:       2356 MB/s (183,308 bytes -> 135,696 bytes)
base64.b64decode:         2156 MB/s (183,308 bytes -> 135,696 bytes)
bench: altchars=None, validate=False, padded=False
pybase64.b64decode:       2332 MB/s (183,308 bytes -> 135,696 bytes)
base64.b64decode:         2079 MB/s (183,308 bytes -> 135,696 bytes)
bench: altchars=b'-_', validate=True, padded=True
pybase64.b64encode:      10328 MB/s (135,696 bytes -> 180,928 bytes)
pybase64.b64decode:       5861 MB/s (180,928 bytes -> 135,696 bytes)
base64.b64encode:         2617 MB/s (135,696 bytes -> 180,928 bytes)
base64.b64decode:         1183 MB/s (180,928 bytes -> 135,696 bytes)
bench: altchars=b'-_', validate=True, padded=False
pybase64.b64encode:      10344 MB/s (135,696 bytes -> 180,928 bytes)
pybase64.b64decode:       2163 MB/s (180,928 bytes -> 135,696 bytes)
base64.b64encode:         2625 MB/s (135,696 bytes -> 180,928 bytes)
base64.b64decode:         1164 MB/s (180,928 bytes -> 135,696 bytes)
bench: altchars=b'-_', ignorechars=b'', padded=True
pybase64.b64decode:       5815 MB/s (180,928 bytes -> 135,696 bytes)
base64.b64decode:         2598 MB/s (180,928 bytes -> 135,696 bytes)
bench: altchars=b'-_', ignorechars=b'', padded=False
pybase64.b64decode:       2146 MB/s (180,928 bytes -> 135,696 bytes)
base64.b64decode:         2586 MB/s (180,928 bytes -> 135,696 bytes)
bench: altchars=b'-_', ignorechars=b'\n', padded=True
pybase64.b64decode:       2042 MB/s (183,308 bytes -> 135,696 bytes)
base64.b64decode:         2056 MB/s (183,308 bytes -> 135,696 bytes)
bench: altchars=b'-_', ignorechars=b'\n', padded=False
pybase64.b64decode:       2034 MB/s (183,308 bytes -> 135,696 bytes)
base64.b64decode:         1952 MB/s (183,308 bytes -> 135,696 bytes)
bench: altchars=b'-_', validate=False, padded=True
pybase64.b64decode:       2021 MB/s (183,308 bytes -> 135,696 bytes)
base64.b64decode:         1063 MB/s (183,308 bytes -> 135,696 bytes)
bench: altchars=b'-_', validate=False, padded=False
pybase64.b64decode:       2014 MB/s (183,308 bytes -> 135,696 bytes)
base64.b64decode:         1042 MB/s (183,308 bytes -> 135,696 bytes)

Changelog

Future

• Speed-up translation on aarch64
• Add padded and wrapcol parameters to b64encode
• Add padded parameter to urlsafe_b64encode
• Add padded parameter to b64decode & urlsafe_b64decode
• urlsafe_b64decode now defaults to padded=False to align with Python 3.15 behavior
• Add ignorechars parameter to b64decode
• Handle excess padding with the same behavior as CPython 3.15
• Reject non-ASCII strings in b64decode when validate=False
• Deprecate accepting the + and / characters with an alternative alphabet when decoding
• Use ValueError instead of AssertionError on altchars length validation
• Add SBOM to PyPI wheels
• Add initial support for Python 3.15
• Drop python 3.8 support
• Stop publishing python 3.13t wheels

1.4.3

• Publish Android Python 3.14 wheels
• Publish GraalPy v25 wheels

1.4.2

• Update base64 library (Windows ARM64 Neon support)
• Publish Python 3.14 wheels
• Publish Linux riscv64 wheels
• Publish Android wheels
• Publish iOS wheels
• Publish GraalPy wheels

1.4.1

• Publish PyPy 3.11 wheels
• Publish armv7l wheels

1.4.0

• Publish python 3.13 wheels
• Add support for free-threaded builds
• Add MSYS2 support for C-extension
• Better logging on base64 build failure when C-extension build is optional
• Drop python 3.6 & 3.7 support

1.3.2

• Update base64 library
• PyPy: fix wrong outcome with non C-contiguous buffer

1.3.1

• Add missing py.typed marker

1.3.0

• Update base64 library
• Add AVX512-VBMI implementation
• Rework extension build to remove adherence on distutils
• Publish python 3.12 wheels
• Documentation now uses furo theme

1.2.3

• Update base64 library
• Publish python 3.11 wheels

1.2.2

• Update base64 library
• Fix C extension build on musl distros
• Publish musllinux wheels

1.2.1

• Publish PyPy 3.8 (pypy38_pp73) wheels

1.2.0

• Release the GIL
• Publish CPython 3.10 wheels
• Drop python 3.5 support

1.1.4

• Add macOS arm64 wheel

1.1.3

• GitHub Actions: fix build on tag

1.1.2

• Add PyPy wheels
• Add aarch64, ppc64le & s390x manylinux wheels

1.1.1

• Move CI from TravisCI/AppVeyor to GitHub Actions
• Fix publication of Linux/macOS wheels

1.1.0

• Add b64encode_as_string, same as b64encode but returns a str object instead of a bytes object
• Add b64decode_as_bytearray, same as b64decode but returns a bytarray object instead of a bytes object
• Speed-Up decoding from UCS1 strings

1.0.2

• Update base64 library
• Publish python 3.9 wheels

1.0.1

• Publish python 3.8 wheels

1.0.0

• Drop python 3.4 support
• Drop python 2.7 support

0.5.0

• Publish python 3.7 wheels
• Drop python 3.3 support

0.4.0

• Speed-up decoding when validate==False

0.3.1

• Fix deployment issues

0.3.0

• Add encodebytes function

0.2.1

• Fixed invalid results on Windows

0.2.0

• Added documentation

• Added subcommands to the main script:

  • help
  • version
  • encode
  • decode
  • benchmark

0.1.2

• Updated base64 native library

0.1.1

• Fixed deployment issues

0.1.0

• First public release