dpnp is published under the BSD-3-Clause license . Its sources state no price, so this page gives none. You can install it with conda, mamba or pip package managers . Check dpnp's own pages for any cost or terms before you depend on it in a project, because the facts used here do not cover them.

dpnp vs MLX in 2026

Q: Is MLX free to use?

MLX is brought to users by Apple machine learning research . No price or licence for MLX is given in the facts used here, so none is published on this page. Check the project's own pages for licence and cost details before you choose it.

Q: Is PyOverdrive free to use?

PyOverdrive is open source under the MIT License . Its claims state no price or fee, so none is published on this page. The version is 1.0.0 and its status is Beta . Because it is Beta, check that status against your project's needs before you rely on it for production work.

dpnp, MLX and PyOverdrive compared on what each is built for, how each is installed and licensed, and what to check before you choose.

By SayDeploy · 2026-10-29

SayDeploy is published by LunarWerx, which makes some of the tools compared here.

This page compares dpnp, MLX and PyOverdrive on what each is built for, what each costs, how each is installed and what to check before you choose one.

Option	Licence	Price or fee	Install	Devices or systems
dpnp	BSD-3-Clause license ^[1]	not published	conda, mamba or pip package managers ^[1]	not published
MLX	not published	not published	not published	CPU and GPU device types ^[2]
PyOverdrive	MIT License ^[3]	not published	pip, straight from the GitHub repository ^[3]	OS independent; the test suite has been run on Windows and Linux x86-64 ^[3]

The short answer

MLX is brought to users by Apple machine learning research ^[2]. It has composable function transformations for automatic differentiation, automatic vectorization and computation graph optimization ^[2].

dpnp's sources here cover only its license and its install routes ^[1] ^[1], so this page says little about what it does.

PyOverdrive's claims say that existing NumPy code can use faster versions of supported functions after one call ^[3].

On needs, MLX's page names automatic differentiation among its transformations ^[2]. The dpnp facts name no need it meets, so check its own documentation before you choose it. PyOverdrive's claims name existing NumPy code as the code it works on ^[3].

What dpnp and MLX are built for

The facts available for dpnp are its license and its install routes, so this section says no more about its purpose. The dpnp repository is published under the BSD-3-Clause license ^[1]. It can be installed using conda, mamba or pip package managers ^[1].

MLX is brought to users by Apple machine learning research ^[2]. It supports composable function transformations for automatic differentiation, automatic vectorization and computation graph optimization ^[2].

What PyOverdrive is built for

PyOverdrive is made by the team that publishes this site.

Calling pyoverdrive.enable() patches supported NumPy functions so your existing code uses faster versions of them, and calls it does not support run on stock NumPy ^[3]. A fast version runs only when a check confirms the input is in its measured range. Anything else, or a fast version that raises an error, goes to stock NumPy ^[3]. python -m pyoverdrive --selfcheck compares every fast version with stock NumPy on your own machine ^[3].

For large float arrays, sin, cos, tan, exp, log, log10 and tanh run on several threads above a size measured for each function and data type, between 3e5 and 3e6 elements ^[3]. On large stacks of 3x3 matrices, np.linalg.qr uses a vectorized closed-form Householder method, benchmarked at 10k matrices ^[3]. Large np.einsum calls are sent through NumPy's own optimize=True planner once they pass a measured size threshold ^[3]. np.isin on 1-D arrays of StringDType or object values can use a hash-based lookup instead of the default method ^[3].

The package is open source under the MIT License ^[3] and is marked OS independent ^[3].

Pricing and fees

dpnp is published under the BSD-3-Clause license ^[1]. Its sources state no price, so none is published on this page.

MLX is brought to users by Apple machine learning research ^[2]. No price for MLX is given in the facts used here, so none is published on this page.

PyOverdrive is open source under the MIT License ^[3]. Its claims state no price or fee, so none is published on this page.

The table above compares the licence, price, install route and device or system details of each option.

Setting it up and daily use

dpnp can be installed with conda, mamba or pip package managers ^[1]. The facts here say nothing about how it is used day to day, so check its own documentation for that.

For MLX, the facts here describe its design rather than its installation. Check its own pages for the install steps. MLX offers composable function transformations for automatic differentiation, automatic vectorization and computation graph optimization ^[2].

You install PyOverdrive with pip straight from the GitHub repository ^[3], and it needs Python 3.12 or newer ^[3]. Calling pyoverdrive.enable() patches supported NumPy functions so existing code uses faster versions, and calls it does not support run on stock NumPy ^[3]. Before you rely on it, python -m pyoverdrive --selfcheck compares every fast version with stock NumPy on your own machine ^[3]. python -m pyoverdrive --calibrate re-times the multi-threaded versions on your CPU and switches off any that do not pay there ^[3]. pyoverdrive.explain() tells you which version a call would take and why, without running the call ^[3]. You can turn off one fast version with disable_path() or PYOVERDRIVE_DISABLE, or call pyoverdrive.disable() to restore the original NumPy functions ^[3].

Things to check before you choose

Ask what your code calls. PyOverdrive's claims cover supported NumPy functions, so compare the functions your code uses with the supported ones ^[3].

Ask whether your system is covered. The table lists the device and system details for each option. PyOverdrive's test suite has been run on Windows and on Linux x86-64 ^[3].

Ask how closely results must match stock NumPy. PyOverdrive's claims say a check against stock NumPy can run on your own machine ^[3].

Ask about maturity. PyOverdrive is marked Beta ^[3], so decide whether that status suits your project.

For dpnp, the facts here are thin, so read its own documentation before you rely on it.

How to choose

If you want faster versions of supported NumPy functions after one call, PyOverdrive's claims describe that ^[3], with stock NumPy as the fallback ^[3].

If you want to see which version a call would take before running it, pyoverdrive.explain() shows that ^[3].

If you want automatic differentiation and function transformations, MLX's page lists them ^[2].

If you want to install dpnp with conda, mamba or pip, its page says those routes exist ^[1].

If you want to check the licence of either tool, dpnp is published under the BSD-3-Clause license ^[1] and PyOverdrive under the MIT License ^[3].

Frequently asked questions

Is dpnp free to use?

dpnp is published under the BSD-3-Clause license ^[1]. Its sources state no price, so this page gives none. You can install it with conda, mamba or pip package managers ^[1]. Check dpnp's own pages for any cost or terms before you depend on it in a project, because the facts used here do not cover them.

Is MLX free to use?

MLX is brought to users by Apple machine learning research ^[2]. No price or licence for MLX is given in the facts used here, so none is published on this page. Check the project's own pages for licence and cost details before you choose it.

Is PyOverdrive free to use?

PyOverdrive is open source under the MIT License ^[3]. Its claims state no price or fee, so none is published on this page. The version is 1.0.0 and its status is Beta ^[3]. Because it is Beta, check that status against your project's needs before you rely on it for production work.

Does PyOverdrive change how existing NumPy code runs?

Calling pyoverdrive.enable() patches supported NumPy functions so existing code uses faster versions of them ^[3]. A fast version runs only when a check confirms the input is in its measured range, and anything else goes to stock NumPy ^[3]. You can turn one fast version off with disable_path(), or call pyoverdrive.disable() to restore the original NumPy functions ^[3].

How do you install PyOverdrive and which Python does it need?

You install PyOverdrive with pip straight from the GitHub repository ^[3]. It needs Python 3.12 or newer ^[3]. The package is marked OS independent, and the test suite has been run on Windows and on Linux x86-64 ^[3]. Run python -m pyoverdrive --demo to time the headline operations with stock NumPy and with PyOverdrive on your machine, in about 20 seconds ^[3].

It is open source under the MIT License. ^[3]

Try PyOverdrive