Key features

• The bit-endianness can be specified for each bitarray object, see below.

• Sequence methods: slicing (including slice assignment and deletion), operations +, *, +=, *=, the in operator, len()

• Bitwise operations: ~, &, |, ^, <<, >> (as well as their in-place versions &=, |=, ^=, <<=, >>=).

• Free-threading support (for GIL-disabled CPython 3.14 and later)

• Written to support all platforms supported by CPython and PyPy.

• Fast methods for encoding and decoding variable-length prefix codes.

• Bitarray objects support the buffer protocol (both importing and exporting buffers).

• Efficient packing and unpacking for interoperability with byte-oriented objects such as NumPy arrays.

• Pickling and unpickling of bitarray objects.

• Immutable frozenbitarray objects which are hashable

• Fast searching for bits and bit patterns

• bitfields sub-module for packing and unpacking fixed-width bit-field structures

• Type hinting

• Extensive test suite with over 700 unit tests

• Utility module bitarray.util:

  • conversion to and from hexadecimal strings

  • generating random bitarrays

  • pretty printing

  • conversion to and from integers

  • creating Huffman codes

  • compression of sparse bitarrays

  • run-length encoding

  • serialization and deserialization

  • various count functions

  • other helpful functions

Installation

Python wheels are are available on PyPI for all major platforms and Python versions. Which means you can simply:

$ python -m pip install bitarray

Once you have installed the package, you may want to test it:

$ python -c 'import bitarray; bitarray.test()'
bitarray is installed in: /Users/ilan/bitarray/bitarray
sys.prefix: /Users/ilan/miniforge
bitarray version: 3.12.3
platform.machine(): arm64
platform.platform(): macOS-14.3-arm64-arm-64bit-Mach-O
sys.platform: darwin
sys.version: 3.15.0 (main, Oct  9 2026) [Clang 21.1.8]
sys.abiflags: ''
sys._is_gil_enabled(): True
pointer size: 64 bit
HAVE_BUILTIN_BSWAP64: 1
default bit-endianness: big
machine byte-order: little
Py_GIL_DISABLED: 0
Py_DEBUG: 0
DEBUG: 0
.........................................................................
................................................s........................
......s.........................................................
----------------------------------------------------------------------
Ran 723 tests in 0.163s

OK (skipped=2)

The test() function is part of the API. It will return a unittest.runner.TextTestResult object, such that one can verify that all tests ran successfully by:

import bitarray
assert bitarray.test().wasSuccessful()

Usage

As mentioned above, bitarray objects behave very much like lists, so there is not too much to learn. The biggest difference from list objects (except that bitarray are obviously homogeneous) is the ability to access the machine representation of the object. When doing so, the bit-endianness is of importance; this issue is explained in detail in the section below. Here, we demonstrate the basic usage of bitarray objects:

>>> from bitarray import bitarray
>>> a = bitarray()         # create empty bitarray
>>> a.append(1)
>>> a.extend([1, 0])
>>> a
bitarray('110')
>>> x = bitarray(2 ** 20)  # bitarray of length 1048576 (initialized to 0)
>>> len(x)
1048576
>>> bitarray('1001 011')   # initialize from string (whitespace is ignored)
bitarray('1001011')
>>> lst = [1, 0, False, True, True]
>>> a = bitarray(lst)      # initialize from iterable
>>> a
bitarray('10011')
>>> a[2]    # indexing a single item will always return an integer
0
>>> a[2:4]  # whereas indexing a slice will always return a bitarray
bitarray('01')
>>> a[2:3]  # even when the slice length is just one
bitarray('0')
>>> a.count(1)
3
>>> a.remove(0)            # removes first occurrence of 0
>>> a
bitarray('1011')

Like lists, bitarray objects support slice assignment and deletion:

>>> a = bitarray(50)
>>> a.setall(0)            # set all elements in a to 0
>>> a[11:37:3] = 9 * bitarray('1')
>>> a
bitarray('00000000000100100100100100100100100100000000000000')
>>> del a[12::3]
>>> a
bitarray('0000000000010101010101010101000000000')
>>> a[-6:] = bitarray('10011')
>>> a
bitarray('000000000001010101010101010100010011')
>>> a += bitarray('000111')
>>> a[9:]
bitarray('001010101010101010100010011000111')

In addition, slices can be assigned to booleans, which is easier (and faster) than assigning to a bitarray in which all values are the same:

>>> a = 20 * bitarray('0')
>>> a[1:15:3] = True
>>> a
bitarray('01001001001001000000')

This is easier and faster than:

>>> a = 20 * bitarray('0')
>>> a[1:15:3] = 5 * bitarray('1')
>>> a
bitarray('01001001001001000000')

Note that in the latter we have to create a temporary bitarray whose length must be known or calculated. Another example of assigning slices to Booleans, is setting ranges:

>>> a = bitarray(30)
>>> a[:] = 0         # set all elements to 0 - equivalent to a.setall(0)
>>> a[10:25] = 1     # set elements in range(10, 25) to 1
>>> a
bitarray('000000000011111111111111100000')

As of bitarray version 2.8, indices may also be lists of arbitrary indices (like in NumPy), or bitarrays that are treated as masks, see Bitarray indexing.

Bitwise operators

Bitarray objects support the bitwise operators ~, &, |, ^, <<, >> (as well as their in-place versions &=, |=, ^=, <<=, >>=). The behavior is very much what one would expect:

>>> a = bitarray('101110001')
>>> ~a  # invert
bitarray('010001110')
>>> b = bitarray('111001011')
>>> a ^ b  # bitwise XOR
bitarray('010111010')
>>> a &= b  # inplace AND
>>> a
bitarray('101000001')
>>> a <<= 2  # in-place left-shift by 2
>>> a
bitarray('100000100')
>>> b >> 1  # return b right-shifted by 1
bitarray('011100101')

The C language does not specify the behavior of negative shifts and of left shifts larger or equal than the width of the promoted left operand. The exact behavior is compiler/machine specific. This Python bitarray library specifies the behavior as follows:

• the length of the bitarray is never changed by any shift operation

• blanks are filled by 0

• negative shifts raise ValueError

• shifts larger or equal to the length of the bitarray result in bitarrays with all values 0

It is worth noting that (regardless of bit-endianness) the bitarray left shift (<<) always shifts towards lower indices, and the right shift (>>) always shifts towards higher indices.

Bit-endianness

For many purposes the bit-endianness is not of any relevance to the end user and can be regarded as an implementation detail of bitarray objects. However, there are use cases when the bit-endianness becomes important. These use cases involve explicitly reading and writing the bitarray buffer using .tobytes(), .frombytes(), .tofile() or .fromfile(), importing and exporting buffers. Also, a number of utility functions in bitarray.util will return different results depending on bit-endianness, such as ba2hex() or ba2int. To better understand this topic, please read bit-endianness.

Buffer protocol

Bitarray objects support the buffer protocol. They can both export their own buffer, as well as import another object’s buffer. To learn more about this topic, please read buffer protocol. There is also an example that shows how to memory-map a file to a bitarray: mmapped-file.py

Frozenbitarrays

A frozenbitarray object is very similar to the bitarray object. The difference is that this a frozenbitarray is immutable, and hashable, and can therefore be used as a dictionary key:

>>> from bitarray import frozenbitarray
>>> key = frozenbitarray('1100011')
>>> {key: 'some value'}
{frozenbitarray('1100011'): 'some value'}
>>> key[3] = 1
Traceback (most recent call last):
    ...
TypeError: frozenbitarray is immutable

The bitarray object:

bitarray(initializer=0, /, endian='big', buffer=None) -> bitarray

Return a new bitarray object whose items are bits initialized from the optional initializer, and bit-endianness. The initializer may be one of the following types: a.) int bitarray, initialized to zeros, of given length b.) bytes or bytearray to initialize buffer directly c.) str of 0s and 1s, ignoring whitespace and “_” d.) iterable of integers 0 or 1.

Optional keyword arguments:

endian: Specifies the bit-endianness of the created bitarray object. Allowed values are big and little (the default is big). The bit-endianness affects the buffer representation of the bitarray.

buffer: Any object which exposes a buffer. When provided, initializer cannot be present (or has to be None). The imported buffer may be read-only or writable, depending on the object type.

New in version 2.3: optional buffer argument

New in version 3.4: allow initializer bytes or bytearray to set buffer directly

Other objects:

frozenbitarray(initializer=0, /, endian='big', buffer=None) -> frozenbitarray

Return a frozenbitarray object. Initialized the same way a bitarray object is initialized. A frozenbitarray is immutable and hashable, and may therefore be used as a dictionary key.

New in version 1.1

decodetree(code, /) -> decodetree

Given a prefix code (a dict mapping symbols to bitarrays), create a binary tree object to be passed to .decode(). The prefix code length cannot exceed 256 bits.

See also: Variable-length prefix codes

New in version 1.6

Functions defined in the bitarray module:

bits2bytes(n, /) -> int

Return the number of bytes necessary to store n bits.

get_default_endian() -> str

Return the default bit-endianness for new bitarray objects being created.

New in version 1.3

test(verbosity=1) -> TextTestResult

Run self-test, and return unittest.runner.TextTestResult object.

Functions defined in bitarray.util module:

This sub-module was added in version 1.2.

any_and(a, b, /) -> bool

Efficient implementation of any(a & b).

New in version 2.7

ba2base(n, bitarray, /, group=0, sep=' ') -> str

Return a string containing the base n ASCII representation of the bitarray. Allowed values for n are 2, 4, 8, 16, 32 and 64. The bitarray has to have a length divisible by 1, 2, 3, 4, 5 or 6 respectively. For n=32 the RFC 4648 Base32 alphabet is used, and for n=64 the standard base 64 alphabet is used. When grouped, the string sep is inserted between groups of group characters, default is a space.

See also: Bitarray representations

New in version 1.9

New in version 3.3: optional group and sep arguments

ba2hex(bitarray, /, group=0, sep=' ') -> hexstr

Return a string containing the hexadecimal representation of the bitarray (which has to be multiple of 4 in length). When grouped, the string sep is inserted between groups of group characters, default is a space.

New in version 3.3: optional group and sep arguments

ba2int(bitarray, /, signed=False) -> int

Convert the given bitarray to an integer. The bit-endianness of the bitarray is respected. signed indicates whether two’s complement is used to represent the integer.

base2ba(n, asciistr, /, endian=None) -> bitarray

Bitarray of base n ASCII representation. Allowed values for n are 2, 4, 8, 16, 32 and 64. For n=32 the RFC 4648 Base32 alphabet is used, and for n=64 the standard base 64 alphabet is used. ASCII whitespace is ignored.

See also: Bitarray representations

New in version 1.9

New in version 3.3: ignore whitespace

byteswap(a, n=<buffer size>, /)

Reverse every n consecutive bytes of a in-place. By default, all bytes are reversed. Note that n is not limited to 2, 4 or 8, but can be any positive integer. Also, a may be any object that exposes a writable buffer. Nothing about this function is specific to bitarray objects.

We should mention that Python’s array.array object has a method .byteswap() with similar functionality. However, unlike bitarray’s util.byteswap() function, this method is limited to swapping 2, 4, or 8 consecutive bytes.

New in version 3.4

canonical_decode(bitarray, count, symbol, /) -> iterator

Decode bitarray using canonical Huffman decoding tables where count is a sequence containing the number of symbols of each length and symbol is a sequence of symbols in canonical order.

See also: Canonical Huffman Coding

New in version 2.5

canonical_huffman(dict, /) -> tuple

Given a frequency map, a dictionary mapping symbols to their frequency, calculate the canonical Huffman code. Returns a tuple containing:

0. the canonical Huffman code as a dict mapping symbols to frozenbitarrays

1. a list containing the number of symbols of each code length

2. a list of symbols in canonical order

Note: the two lists may be used as input for canonical_decode().

See also: Canonical Huffman Coding

New in version 2.5

New in version 3.10: return a codedict mapping to frozenbitarrays

correspond_all(a, b, /) -> tuple

Return tuple with counts of: ~a & ~b, ~a & b, a & ~b, a & b

New in version 3.4

count_and(a, b, /) -> int

Return (a & b).count() in a memory efficient manner, as no intermediate bitarray object gets created.

count_n(a, n, value=1, /) -> int

Return lowest index i for which a[:i].count(value) == n. Raises ValueError when n exceeds total count (a.count(value)).

New in version 2.3.6: optional value argument

count_or(a, b, /) -> int

Return (a | b).count() in a memory efficient manner, as no intermediate bitarray object gets created.

count_xor(a, b, /) -> int

Return (a ^ b).count() in a memory efficient manner, as no intermediate bitarray object gets created.

This is also known as the Hamming distance.

deserialize(bytes, /) -> bitarray

Return a bitarray given a bytes-like representation such as returned by serialize().

See also: Bitarray representations

New in version 1.8

New in version 2.5.0: allow bytes-like argument

gen_primes(n, /, endian=None, odd=False) -> bitarray

Generate a bitarray of length n in which active indices are prime numbers. By default (odd=False), active indices correspond to prime numbers directly. When odd=True, only odd prime numbers are represented in the resulting bitarray a, and a[i] corresponds to 2*i+1 being prime or not.

Apart from working with prime numbers, this function is useful for testing, as it provides a simple way to create a well-defined bitarray of any length.

New in version 3.7

hex2ba(hexstr, /, endian=None) -> bitarray

Bitarray of hexadecimal representation. hexstr may contain any number (including odd numbers) of hex digits (upper or lower case). ASCII whitespace is ignored.

New in version 3.3: ignore whitespace

huffman_code(dict, /, endian=None) -> dict

Given a frequency map, a dictionary mapping symbols to their frequency, calculate the Huffman code, i.e. a dict mapping those symbols to frozenbitarrays (with given bit-endianness). Note that the symbols are not limited to being strings. Symbols may be any hashable object.

See also: Variable-length prefix codes

New in version 3.10: return a codedict mapping to frozenbitarrays

int2ba(int, /, length=None, endian=None, signed=False) -> bitarray

Convert the given integer to a bitarray (with given bit-endianness, and no leading (big-endian) / trailing (little-endian) zeros), unless the length of the bitarray is provided. An OverflowError is raised if the integer is not representable with the given number of bits. signed determines whether two’s complement is used to represent the integer, and requires length to be provided.

intervals(bitarray, /) -> iterator

Compute all uninterrupted intervals of 1s and 0s, and return an iterator over tuples (value, start, stop). The intervals are guaranteed to be in order, and their size is always non-zero (stop - start > 0).

New in version 2.7

ones(n, /, endian=None) -> bitarray

Create a bitarray of length n, with all values 1, and optional bit-endianness (little or big).

New in version 2.9

parity(a, /) -> int

Return parity of bitarray a. parity(a) is equivalent to a.count() % 2 but more efficient.

New in version 1.9

pprint(bitarray, /, stream=None, group=8, indent=4, width=80)

Pretty-print bitarray object to stream, defaults to sys.stdout. By default, bits are grouped in bytes (8 bits), and 64 bits per line. Non-bitarray objects are printed using pprint.pprint().

New in version 1.8

random_k(n, /, k, endian=None) -> bitarray

Return (pseudo-) random bitarray of length n with k elements set to one. Mathematically equivalent to setting (in a bitarray of length n) all bits at indices random.sample(range(n), k) to one. The random bitarrays are reproducible when calling Python’s random.seed() with a specific seed value.

New in version 3.6

random_p(n, /, p=0.5, endian=None) -> bitarray

Return (pseudo-) random bitarray of length n, where each bit has probability p of being one (independent of any other bits). Mathematically equivalent to bitarray((random() < p for _ in range(n)), endian), but much faster for large n. The random bitarrays are reproducible when calling Python’s random.seed() with a specific seed value.

This function requires Python 3.12 or higher, as it depends on the standard library function random.binomialvariate(). Raises NotImplementedError when Python version is too low.

See also: Random Bitarrays

New in version 3.5

rl_decode(stream, /, endian=None) -> bitarray

Decode a run-length encoded bitarray from a binary stream (an integer iterator, or bytes-like object), and return the decoded bitarray. This function consumes only one bitarray and leaves the remaining stream untouched. Use rl_encode() for encoding.

See also: Comparison of bitarray codecs

New in version 3.11

rl_encode(bitarray, /) -> bytes

Return the run-length encoded binary representation of a bitarray. This representation may be useful for bitarrays with long runs of identical bits. Use rl_decode() for decoding.

See also: Comparison of bitarray codecs

New in version 3.11

sc_decode(stream, /) -> bitarray

Decompress binary stream (an integer iterator, or bytes-like object) of a sparse compressed (sc) bitarray, and return the decoded bitarray. This function consumes only one bitarray and leaves the remaining stream untouched. Use sc_encode() for compressing (encoding).

See also: Compression of sparse bitarrays, Comparison of bitarray codecs

New in version 2.7

sc_encode(bitarray, /) -> bytes

Compress a bitarray using sparse encoding and return its binary representation. This representation is useful for efficiently storing sparse bitarrays. Use sc_decode() for decompressing (decoding).

See also: Compression of sparse bitarrays, Comparison of bitarray codecs

New in version 2.7

sc_stat(stream) -> dict

Scan a stream produced by sc_encode() and return a dictionary containing its bit-endianness (endian), length (nbits), and the number of blocks of each type (blocks). blocks is a list such that blocks[i] is the number of blocks of type i.

Like sc_decode(), it consumes one encoded bitarray and leaves remaining input untouched.

See also: Compression of sparse bitarrays

New in version 3.11

serialize(bitarray, /) -> bytes

Return a serialized representation of the bitarray, which may be passed to deserialize(). It efficiently represents the bitarray object (including its bit-endianness) and is guaranteed not to change in future releases.

See also: Bitarray representations

New in version 1.8

strip(bitarray, /, mode='right') -> bitarray

Return a new bitarray with zeros stripped from left, right or both ends. Allowed values for mode are the strings: left, right, both

subset(a, b, /) -> bool

Return True if bitarray a is a subset of bitarray b. subset(a, b) is equivalent to a | b == b (and equally a & b == a) but more efficient as no intermediate bitarray object is created and the buffer iteration is stopped as soon as one mismatch is found.

sum_indices(a, /, mode=1) -> int

Return sum of indices of all active bits in bitarray a. Equivalent to sum(i for i, v in enumerate(a) if v). mode=2 sums square of indices.

New in version 3.6

New in version 3.7: add optional mode argument

urandom(n, /, endian=None) -> bitarray

Return random bitarray of length n (uses os.urandom()).

New in version 1.7

vl_decode(stream, /, endian=None) -> bitarray

Decode binary stream (an integer iterator, or bytes-like object), and return the decoded bitarray. This function consumes only one bitarray and leaves the remaining stream untouched. Use vl_encode() for encoding.

See also: Variable length bitarray format, Comparison of bitarray codecs

New in version 2.2

vl_encode(bitarray, /) -> bytes

Return variable length binary representation of bitarray. This representation is useful for efficiently storing small bitarray in a binary stream. Use vl_decode() for decoding.

See also: Variable length bitarray format, Comparison of bitarray codecs

New in version 2.2

xor_indices(a, /) -> int

Return xor reduced indices of all active bits in bitarray a. This is essentially equivalent to reduce(operator.xor, (i for i, v in enumerate(a) if v)).

New in version 3.2

zeros(n, /, endian=None) -> bitarray

Create a bitarray of length n, with all values 0, and optional bit-endianness (little or big).

The bitarray.bitfields module:

This sub-module was added in version 3.12. For a detailed description, see: Bit-field structures

Struct(format) -> compiled struct object

Central class for packing and unpacking bit-level structures.

calcsize(format) -> int

Return the size (in bits) of the struct corresponding to the format string.

compile(format) -> Struct

Compile given format string and return a compiled format object that can be used to pack and/or unpack data multiple times.

pack(format, v1, v2, ...) -> bitarray

Return a bitarray containing the values v1, v2, … packed according to the format string.

pack_into(format, bitarray, offset, v1, v2, ...) -> None

Pack the values v1, v2, … into the writable bitarray starting at bit offset offset. A negative offset counts from the end of bitarray.

unpack(format, bitarray) -> tuple

Return a tuple containing values unpacked according to the format string.

unpack_from(format, bitarray, offset=0) -> tuple

Unpack values from bitarray starting at bit offset offset, and return a tuple. A negative offset counts from the end of bitarray.