Installing NumPy

The only prerequisite for NumPy is Python itself. If you don’t have Python yet and want the simplest way to get started, we recommend you use the Anaconda Distribution - it includes Python, NumPy, and other commonly used packages for scientific computing and data science.

NumPy can be installed with conda, with pip, or with a package manager on macOS and Linux. For more detailed instructions, consult our Python and NumPy installation guide below.


If you use conda, you can install it with:

conda install numpy


If you use pip, you can install it with:

pip install numpy

Python and NumPy installation guide

Installing and managing packages in Python is complicated, there are a number of alternative solutions for most tasks. This guide tries to give the reader a sense of the best (or most popular) solutions, and give clear recommendations. It focuses on users of Python, NumPy, and the PyData (or numerical computing) stack on common operating systems and hardware.


We’ll start with recommendations based on the user’s experience level and operating system of interest. If you’re in between “beginning” and “advanced”, please go with “beginning” if you want to keep things simple, and with “advanced” if you want to work according to best practices that go a longer way in the future.

Beginning users

On all of Windows, macOS, and Linux:

Advanced users

Windows or macOS


If you’re fine with slightly outdated packages and prefer stability over being able to use the latest versions of libraries:

If you use a GPU:


Alternative if you prefer pip/PyPI

For users who know, from personal preference or reading about the main differences between conda and pip below, they prefer a pip/PyPI-based solution, we recommend:

Python package management

Managing packages is a challenging problem, and, as a result, there are lots of tools. For web and general purpose Python development there’s a whole host of tools complementary with pip. For high-performance computing (HPC), Spack is worth considering. For most NumPy users though, conda and pip are the two most popular tools.

Pip & conda

The two main tools that install Python packages are pip and conda. Their functionality partially overlaps (e.g. both can install numpy), however, they can also work together. We’ll discuss the major differences between pip and conda here - this is important to understand if you want to manage packages effectively.

The first difference is that conda is cross-language and it can install Python, while pip is installed for a particular Python on your system and installs other packages to that same Python install only. This also means conda can install non-Python libraries and tools you may need (e.g. compilers, CUDA, HDF5), while pip can’t.

The second difference is that pip installs from the Python Packaging Index (PyPI), while conda installs from its own channels (typically “defaults” or “conda-forge”). PyPI is the largest collection of packages by far, however, all popular packages are available for conda as well.

The third difference is that pip does not have a dependency resolver (this is expected to change in the near future), while conda does. For simple cases (e.g. you just want NumPy, SciPy, Matplotlib, Pandas, Scikit-learn, and a few other packages) that doesn’t matter, however, for complicated cases conda can be expected to do a better job keeping everything working well together. The flip side of that coin is that installing with pip is typically a lot faster than installing with conda.

The fourth difference is that conda is an integrated solution for managing packages, dependencies and environments, while with pip you may need another tool (there are many!) for dealing with environments or complex dependencies.

Reproducible installs

Making the installation of all the packages your analysis, library or application depends on reproducible is important. Sounds obvious, yet most users don’t think about doing this (at least until it’s too late).

The problem with Python packaging is that sooner or later, something will break. It’s not often this bad,

Python Environment XKCD image

XKCD illustration - Python environment degradation

but it does degrade over time. Hence, it’s important to be able to delete and reconstruct the set of packages you have installed.

Best practice is to use a different environment per project you’re working on, and record at least the names (and preferably versions) of the packages you directly depend on in a static metadata file. Each packaging tool has its own metadata format for this:

Sometimes it’s too much overhead to create and switch between new environments for small tasks. In that case we encourage you to not install too many packages into your base environment, and keep track of versions of packages some other way (e.g. comments inside files, or printing numpy.__version__ after importing it in notebooks).

NumPy packages & accelerated linear algebra libraries

NumPy doesn’t depend on any other Python packages, however, it does depend on an accelerated linear algebra library - typically Intel MKL or OpenBLAS. Users don’t have to worry about installing those, but it may still be important to understand how the packaging is done and how it affects performance and behavior users see.

The NumPy wheels on PyPI, which is what pip installs, are built with OpenBLAS. The OpenBLAS libraries are shipped within the wheels itself. This makes those wheels larger, and if a user installs (for example) SciPy as well, they will now have two copies of OpenBLAS on disk.

In the conda defaults channel, NumPy is built against Intel MKL. MKL is a separate package that will be installed in the users’ environment when they install NumPy. That MKL package is a lot larger than OpenBLAS, several hundred MB. MKL is typically a little faster and more robust than OpenBLAS.

In the conda-forge channel, NumPy is built against a dummy “BLAS” package. When a user installs NumPy from conda-forge, that BLAS package then gets installed together with the actual library - this defaults to OpenBLAS, but it can also be MKL (from the defaults channel), or even BLIS or reference BLAS.

Besides install sizes, performance and robustness, there are two more things to consider: