Release v2.3.0 (June 7, 2025) · NumPy/NumPy

The NumPy 2.3.0 release continues the work to improve free threaded
Python support and annotations together with the usual set of bug fixes.
It is unusual in the number of expired deprecations, code
modernizations, and style cleanups. The latter may not be visible to
users, but is important for code maintenance over the long term. Note
that we have also upgraded from manylinux2014 to manylinux_2_28.

Users running on a Mac having an M4 cpu might see various warnings about
invalid values and such. The warnings are a known problem with
Accelerate. They are annoying, but otherwise harmless. Apple promises to
fix them.

This release supports Python versions 3.11-3.13, Python 3.14 will be
supported when it is released.

Highlights

• Interactive examples in the NumPy documentation.
• Building NumPy with OpenMP Parallelization.
• Preliminary support for Windows on ARM.
• Improved support for free threaded Python.
• Improved annotations.

New functions

New function numpy.strings.slice

The new function numpy.strings.slice was added, which implements fast
native slicing of string arrays. It supports the full slicing API
including negative slice offsets and steps.

(gh-27789)

Deprecations

• The numpy.typing.mypy_plugin has been deprecated in favor of
  platform-agnostic static type inference. Please remove
  numpy.typing.mypy_plugin from the plugins section of your mypy
  configuration. If this change results in new errors being reported,
  kindly open an issue.

  (gh-28129)

• The numpy.typing.NBitBase type has been deprecated and will be
  removed in a future version.

  This type was previously intended to be used as a generic upper
  bound for type-parameters, for example:

  import numpy as np import numpy.typing as npt def f[NT: npt.NBitBase](x: np.complexfloating[NT]) -> np.floating[NT]: ...

  But in NumPy 2.2.0, float64 and complex128 were changed to
  concrete subtypes, causing static type-checkers to reject
  x: np.float64 = f(np.complex128(42j)).

  So instead, the better approach is to use typing.overload:

  import numpy as np from typing import overload @overload def f(x: np.complex64) -> np.float32: ... @overload def f(x: np.complex128) -> np.float64: ... @overload def f(x: np.clongdouble) -> np.longdouble: ...

  (gh-28884)

Expired deprecations

• Remove deprecated macros like NPY_OWNDATA from Cython interfaces
  in favor of NPY_ARRAY_OWNDATA (deprecated since 1.7)

  (gh-28254)

• Remove numpy/npy_1_7_deprecated_api.h and C macros like
  NPY_OWNDATA in favor of NPY_ARRAY_OWNDATA (deprecated since 1.7)

  (gh-28254)

• Remove alias generate_divbyzero_error to
  npy_set_floatstatus_divbyzero and generate_overflow_error to
  npy_set_floatstatus_overflow (deprecated since 1.10)

  (gh-28254)

• Remove np.tostring (deprecated since 1.19)

  (gh-28254)

• Raise on np.conjugate of non-numeric types (deprecated since 1.13)

  (gh-28254)

• Raise when using np.bincount(...minlength=None), use 0 instead
  (deprecated since 1.14)

  (gh-28254)

• Passing shape=None to functions with a non-optional shape argument
  errors, use () instead (deprecated since 1.20)

  (gh-28254)

• Inexact matches for mode and searchside raise (deprecated since
  1.20)

  (gh-28254)

• Setting __array_finalize__ = None errors (deprecated since 1.23)

  (gh-28254)

• np.fromfile and np.fromstring error on bad data, previously they
  would guess (deprecated since 1.18)

  (gh-28254)

• datetime64 and timedelta64 construction with a tuple no longer
  accepts an event value, either use a two-tuple of (unit, num) or a
  4-tuple of (unit, num, den, 1) (deprecated since 1.14)

  (gh-28254)

• When constructing a dtype from a class with a dtype attribute,
  that attribute must be a dtype-instance rather than a thing that can
  be parsed as a dtype instance (deprecated in 1.19). At some point
  the whole construct of using a dtype attribute will be deprecated
  (see #25306)

  (gh-28254)

• Passing booleans as partition index errors (deprecated since 1.23)

  (gh-28254)

• Out-of-bounds indexes error even on empty arrays (deprecated since
  1.20)

  (gh-28254)

• np.tostring has been removed, use tobytes instead (deprecated
  since 1.19)

  (gh-28254)

• Disallow make a non-writeable array writeable for arrays with a base
  that do not own their data (deprecated since 1.17)

  (gh-28254)

• concatenate() with axis=None uses same-kind casting by
  default, not unsafe (deprecated since 1.20)

  (gh-28254)

• Unpickling a scalar with object dtype errors (deprecated since 1.20)

  (gh-28254)

• The binary mode of fromstring now errors, use frombuffer instead
  (deprecated since 1.14)

  (gh-28254)

• Converting np.inexact or np.floating to a dtype errors
  (deprecated since 1.19)

  (gh-28254)

• Converting np.complex, np.integer, np.signedinteger,
  np.unsignedinteger, np.generic to a dtype errors (deprecated
  since 1.19)

  (gh-28254)

• The Python built-in round errors for complex scalars. Use
  np.round or scalar.round instead (deprecated since 1.19)

  (gh-28254)

• 'np.bool' scalars can no longer be interpreted as an index
  (deprecated since 1.19)

  (gh-28254)

• Parsing an integer via a float string is no longer supported.
  (deprecated since 1.23) To avoid this error you can

  • make sure the original data is stored as integers.
  • use the converters=float keyword argument.
  • Use np.loadtxt(...).astype(np.int64)

  (gh-28254)

• The use of a length 1 tuple for the ufunc signature errors. Use
  dtype or fill the tuple with None (deprecated since 1.19)

  (gh-28254)

• Special handling of matrix is in np.outer is removed. Convert to a
  ndarray via matrix.A (deprecated since 1.20)

  (gh-28254)

• Removed the np.compat package source code (removed in 2.0)

  (gh-28961)

C API changes

• NpyIter_GetTransferFlags is now available to check if the iterator
  needs the Python API or if casts may cause floating point errors
  (FPE). FPEs can for example be set when casting float64(1e300) to
  float32 (overflow to infinity) or a NaN to an integer (invalid
  value).

  (gh-27883)

• NpyIter now has no limit on the number of operands it supports.

  (gh-28080)

New NpyIter_GetTransferFlags and NpyIter_IterationNeedsAPI change

NumPy now has the new NpyIter_GetTransferFlags function as a more
precise way checking of iterator/buffering needs. I.e. whether the
Python API/GIL is required or floating point errors may occur. This
function is also faster if you already know your needs without
buffering.

The NpyIter_IterationNeedsAPI function now performs all the checks
that were previously performed at setup time. While it was never
necessary to call it multiple times, doing so will now have a larger
cost.

(gh-27998)

New Features

• The type parameter of np.dtype now defaults to typing.Any. This
  way, static type-checkers will infer dtype: np.dtype as
  dtype: np.dtype[Any], without reporting an error.

  (gh-28669)

• Static type-checkers now interpret:

  • _: np.ndarray as _: npt.NDArray[typing.Any].
  • _: np.flatiter as _: np.flatiter[np.ndarray].

  This is because their type parameters now have default values.

  (gh-28940)

NumPy now registers its pkg-config paths with the pkgconf PyPI package

The pkgconf PyPI
package provides an interface for projects like NumPy to register their
own paths to be added to the pkg-config search path. This means that
when using pkgconf
from PyPI, NumPy will be discoverable without needing for any custom
environment configuration.

(gh-28214)

Allow out=... in ufuncs to ensure array result

NumPy has the sometimes difficult behavior that it currently usually
returns scalars rather than 0-D arrays (even if the inputs were 0-D
arrays). This is especially problematic for non-numerical dtypes (e.g.
object).

For ufuncs (i.e. most simple math functions) it is now possible to use
out=... (literally `...`, e.g. out=Ellipsis) which is identical
in behavior to out not being passed, but will ensure a non-scalar
return. This spelling is borrowed from arr1d[0, ...] where the ...
also ensures a non-scalar return.

Other functions with an out= kwarg should gain support eventually.
Downstream libraries that interoperate via __array_ufunc__ or
__array_function__ may need to adapt to support this.

(gh-28576)

Building NumPy with OpenMP Parallelization

NumPy now supports OpenMP parallel processing capabilities when built
with the -Denable_openmp=true Meson build flag. This feature is
disabled by default. When enabled, np.sort and np.argsort functions
can utilize OpenMP for parallel thread execution, improving performance
for these operations.

(gh-28619)

Interactive examples in the NumPy documentation

The NumPy documentation includes a number of examples that can now be
run interactively in your browser using WebAssembly and Pyodide.

Please note that the examples are currently experimental in nature and
may not work as expected for all methods in the public API.

(gh-26745)

Improvements

• Scalar comparisons between non-comparable dtypes such as
  np.array(1) == np.array('s') now return a NumPy bool instead of a
  Python bool.

  (gh-27288)

• np.nditer now has no limit on the number of supported operands
  (C-integer).

  (gh-28080)

• No-copy pickling is now supported for any array that can be
  transposed to a C-contiguous array.

  (gh-28105)

• The __repr__ for user-defined dtypes now prefers the __name__ of
  the custom dtype over a more generic name constructed from its
  kind and itemsize.

  (gh-28250)

• np.dot now reports floating point exceptions.

  (gh-28442)

• np.dtypes.StringDType is now a generic
  type which
  accepts a type argument for na_object that defaults to
  typing.Never. For example, StringDType(na_object=None) returns a
  StringDType[None], and StringDType() returns a
  StringDType[typing.Never].

  (gh-28856)

Added warnings to np.isclose

Added warning messages if at least one of atol or rtol are either
np.nan or np.inf within np.isclose.

• Warnings follow the user's np.seterr settings

(gh-28205)

Performance improvements and changes

Performance improvements to np.unique

np.unique now tries to use a hash table to find unique values instead
of sorting values before finding unique values. This is limited to
certain dtypes for now, and the function is now faster for those dtypes.
The function now also exposes a sorted parameter to allow returning
unique values as they were found, instead of sorting them afterwards.

(gh-26018)

Performance improvements to np.sort and np.argsort

np.sort and np.argsort functions now can leverage OpenMP for
parallel thread execution, resulting in up to 3.5x speedups on x86
architectures with AVX2 or AVX-512 instructions. This opt-in feature
requires NumPy to be built with the -Denable_openmp Meson flag. Users
can control the number of threads used by setting the OMP_NUM_THREADS
environment variable.

(gh-28619)

Performance improvements for np.float16 casts

Earlier, floating point casts to and from np.float16 types were
emulated in software on all platforms.

Now, on ARM devices that support Neon float16 intrinsics (such as recent
Apple Silicon), the native float16 path is used to achieve the best
performance.

(gh-28769)

Changes

• The vector norm ord=inf and the matrix norms
  ord={1, 2, inf, 'nuc'} now always returns zero for empty arrays.
  Empty arrays have at least one axis of size zero. This affects
  np.linalg.norm, np.linalg.vector_norm, and
  np.linalg.matrix_norm. Previously, NumPy would raises errors or
  return zero depending on the shape of the array.

  (gh-28343)

• A spelling error in the error message returned when converting a
  string to a float with the method np.format_float_positional has
  been fixed.

  (gh-28569)

• NumPy's __array_api_version__ was upgraded from 2023.12 to
  2024.12.

• numpy.count_nonzero for axis=None (default) now returns a NumPy
  scalar instead of a Python integer.

• The parameter axis in numpy.take_along_axis function has now a
  default value of -1.

  (gh-28615)

• Printing of np.float16 and np.float32 scalars and arrays have
  been improved by adjusting the transition to scientific notation
  based on the floating point precision. A new legacy
  np.printoptions mode '2.2' has been added for backwards
  compatibility.

  (gh-28703)

• Multiplication between a string and integer now raises OverflowError
  instead of MemoryError if the result of the multiplication would
  create a string that is too large to be represented. This follows
  Python's behavior.

  (gh-29060)

unique_values may return unsorted data

The relatively new function (added in NumPy 2.0) unique_values may now
return unsorted results. Just as unique_counts and unique_all these
never guaranteed a sorted result, however, the result was sorted until
now. In cases where these do return a sorted result, this may change in
future releases to improve performance.

(gh-26018)

Changes to the main iterator and potential numerical changes

The main iterator, used in math functions and via np.nditer from
Python and NpyIter in C, now behaves differently for some buffered
iterations. This means that:

• The buffer size used will often be smaller than the maximum buffer
  sized allowed by the buffersize parameter.
• The "growinner" flag is now honored with buffered reductions when
  no operand requires buffering.

For np.sum() such changes in buffersize may slightly change numerical
results of floating point operations. Users who use "growinner" for
custom reductions could notice changes in precision (for example, in
NumPy we removed it from einsum to avoid most precision changes and
improve precision for some 64bit floating point inputs).

(gh-27883)

The minimum supported GCC version is now 9.3.0

The minimum supported version was updated from 8.4.0 to 9.3.0, primarily
in order to reduce the chance of platform-specific bugs in old GCC
versions from causing issues.

(gh-28102)

Changes to automatic bin selection in numpy.histogram

The automatic bin selection algorithm in numpy.histogram has been
modified to avoid out-of-memory errors for samples with low variation.
For full control over the selected bins the user can use set the bin
or range parameters of numpy.histogram.

(gh-28426)

Build manylinux_2_28 wheels

Wheels for linux systems will use the manylinux_2_28 tag (instead of
the manylinux2014 tag), which means dropping support for
redhat7/centos7, amazonlinux2, debian9, ubuntu18.04, and other
pre-glibc2.28 operating system versions, as per the PEP 600 support
table.

(gh-28436)

Remove use of -Wl,-ld_classic on macOS

Remove use of -Wl,-ld_classic on macOS. This hack is no longer needed by
Spack, and results in libraries that cannot link to other libraries
built with ld (new).

(gh-28713)

Re-enable overriding functions in the numpy.strings

Re-enable overriding functions in the numpy.strings module.

(gh-28741)

Checksums

MD5

SHA256