Version 0.63.0 (8 December 2025)

This is a major Numba release. Numba now supports Python 3.14. This release also includes initial support for free-threading enabled Python builds.

Intel x86_64 macOS (osx-64) support has been deprecated and moved to Tier 2 support. See Support Tiers for more details.

Please find a summary of all noteworthy items below.

New Features

Initial support for free-threading enabled Python

The Numba code base now supports the CPython interpreter built with free-threading enabled. Note that this feature is experimental and subject to further development.

Also note that the deprecated numba.pycc module does not support the creation of free-threaded AOT compiled modules.

(PR-#10226)

Add support for math.exp2.

Support for math.exp2 is added.

(PR-#10276)

Improvements

Support for Python 3.14

Support for Python 3.14 is implemented.

(PR-#10150)

CUDA API Changes

Fix CUDA target crash bug with NVIDIA driver version 580

Fixes a bug in the CUDA driver wrapper that caused the CUDA target to crash with NVIDIA driver version 580 and later. This was already fixed in the numba-cuda package, but has been backported here for users who haven’t migrated yet. (We still encourage all CUDA users to switch to the numba-cuda package for the most up to date CUDA support in Numba.)

(PR-#10237)

Bug Fixes

Fix the caching of functions that call parallel=True functions.

Previously, reload_init, which keeps track of functions that need to be called when reloading a module, was only maintained for the current function but not called functions. So, if the called function needed something reloaded and the parent function was cached then the reload would not happen. This has been fixed by propagating the reload_init up the call chain.

(PR-#9950)

Make types.ListType string representation evaluable

Fix types.ListType canonical string representation by making it evaluable. This changes str() to repr() call for obtaining string representation of the list’s element.

(PR-#10195)

Fix UnboundLocalError when calling copy on non-Array types

Enables calling np.copy inside a jitted function with non-Array arguments, such as lists or scalars, and fails with a readable error message on incoercible types

(PR-#10201)

Fix np.ndenumerate support for zero-dimensional arrays

np.ndenumerate() now correctly handles zero-dimensional arrays in JIT-compiled functions. Previously this would raise an IndexError.

(PR-#10288)

Fix gufunc memory leak when called from JIT-compiled functions

Fix memory leak in gufuncs when called from within @jit functions. The issue was caused by incorrect reference counting in the gufunc loopnest that increment reference count of array slices.

(PR-#10298)

Deprecations

Deprecation of macOS x86-64 (Intel) support

Official support for the Intel x86_64 macOS platform (also known as osx-64) has been deprecated and moved to Tier 2 under the Support Tiers policy. Continuous integration will no longer test this platform and new releases will not include binary artifacts for osx-64. The platform remains in the source tree and may continue to work based on community reports.

Users on Intel Macs should use the last officially supported release (0.62.x). See https://github.com/numba/numba/issues/10187 for more details.

Documentation Changes

Document the Numba organisation hardware/OS/packaging etc. support policy.

Documentation is added to provide clarity over what OS/hardware/packaging etc. is supported by projects in the Numba organisation.

(PR-#10283)

Pull-Requests