Software downloads

This page provides quick access to various Coral software. However, if you're just getting started, you should follow the appropriate setup guide for your device.

Jump to a section:

Note: The Edge TPU Compiler is available only for Linux, in a Debian package.

Mendel Linux link

The following packages include the Mendel system image for the Coral Dev Board. It also includes a recovery.img file, which you can flash to an SD card and recover a board that fails to boot.

System requirements:

For instructions, see how to flash a system image.

Release Package Size SHA-256 checksum
5.0 Eagle (Jul 2020) 547 MB enterprise-eagle-20200724205123.sha256
4.0 Day (Nov 2019) 541 MB a21205914e46a75ca5bdc9733260a8b6dc4c001513556fd896089bdcd82b596b
3.0 Chef (Apr 2019) 454 MB df733fd22de156f324a1b5fc3d03251a732ddbbf0e34bd54286d42a2e79246e5
2.0 Beaker (Mar 2019) 1.3 GB 85a1db9a6d251a38f34fabf808b4ad3c35d7ab413318bee1d70de48bde776486

Also see the Mendel release notes.

Debian packages link

All Coral software tools and libraries are available as Debian packages. If you're using a Debian system (including Mendel on a Coral Dev Board), we recommend using these packages instead of the other downloads on this page.

You can install and upgrade the packages from the table below using apt-get, but you must first add our package repo to your sources list as follows:

echo "deb coral-edgetpu-stable main" | sudo tee /etc/apt/sources.list.d/coral-edgetpu.list

# This key is pre-installed on the Coral Dev Board
curl | sudo apt-key add -

sudo apt-get update

Then you can install packages from the below table like this:

sudo apt-get install <package_name>
Package nameDescriptionSupported systems1
edgetpu-compiler The Edge TPU Compiler.
  • Debian 10+
  • x86-64
libedgetpu1-max The Edge TPU runtime. (Pre-installed on the Dev Board.) The separate versions affect only USB-based devices to determine the operating frequency: libedgetpu1-max runs at the maximum frequency (500 MHz) and libedgetpu1-std runs at the reduced frequency (250 MHz).

For PCIe-based devices, both packages behave the same because the PCIe driver performs dynamic frequency scaling to adjust the operating frequency based on programmable temperature thresholds.

  • Debian 10+, Ubuntu 16.04+, or
    Raspberry Pi OS Buster and newer
  • x86-64 or Armv8 (64-bit)
libedgetpu-dev The edgetpu_c.h and edgetpu.h header files for running inference with C++.
gasket-dkms Coral driver for PCIe-based Edge TPU devices, such as the M.2 and Mini PCIe Accelerator.
  • Debian 10+ or Ubuntu 16.04+
  • x86-64 or Armv8 (64-bit)
python3-edgetpu The Edge TPU Python API. (Pre-installed on the Dev Board.)

  • Debian 10+, Ubuntu 16.04+, or
    Raspberry Pi OS Buster and newer
  • x86-64, Armv7 (32-bit), or Armv8 (64-bit)
  • Python 3.6, 3.7, or 3.8
edgetpu-examples Code examples for the Edge TPU Python API. Saved in /usr/share/edgetpu/examples/.

python3-coral-enviro The Coral Environmental Board API.

  • Python 3.0+

1 The "supported systems" are system configurations we currently test, but the packages might work on other similar systems. You can also compile the Edge TPU runtime and Python library for your specific platform using our source code.

Edge TPU runtime link

The following ZIPs include the Edge TPU runtime for macOS and Windows, plus the USB and PCIe drivers required on Windows.

System requirements:

  • Windows 10 or macOS 10.15 Catalina
  • x86-64 system architecture

If you're using Debian Linux (including Mendel), you should instead install the Edge TPU runtime from our Debian packages.

Note: We periodically update the Edge TPU runtime with small changes such as to improve support for different host platforms, but the underlying runtime behavior remains the same. That's why the following table shows multiple ZIP packages with the same runtime version and different dates in the filename.

You should always use the latest runtime, and be sure your models are compiled with the corresponding version of the Edge TPU Compiler.

Version Package Size SHA-256 checksum
16.8 MB f5e1dfd26c37641a4c8eca61f9236e31c355302e5a75c81c690626e777fff67a
16.6 MB 8742039d19715c3274152fe144f1be1f0b3f1dda2e5d605ce90cdac47fc95b28
16.6 MB f98857a43f718d129dd6c1565358a71c8e8008d7947332b488c160c756d5d4ad
12 edgetpu_runtime_20190920.tar.gz
2.8 MB 111c1cef3d8079f5ee9ac8c3cd3ec0822f4a784bc8c8aa4a4affe67d3a6b72a4

Also see the Edge TPU runtime release notes.

Edge TPU Python API link

The following Python wheels are for the Edge TPU Python API. If you instead want the TensorFlow Lite API, you can get that as a Python wheel from the TensorFlow Lite Python quickstart.

If you're using Debian Linux (including Mendel), you should instead install the python3-edgetpu Debian package.

To install, run pip3 install and pass it a Python wheel URL from the following table. For example, if you're on a Mac with Python 3.7, install the Edge TPU Python API as follows:

pip3 install

This table lists only the latest version of the Edge TPU Python library.

macOS 10.15 3.5
Windows 10 3.5
Generic Linux 3.X

Pre-compiled models link

For demonstration and experimentation purposes, you can download and run one of several pre-compiled models that are compatible with the Edge TPU.

Legacy packages link

The following packages are here for archival purposes only.

Release Package Size SHA-256 checksum
2.12.1 (Sep 2019) edgetpu_runtime_20190920.tar.gz
2.8 MB 111c1cef3d8079f5ee9ac8c3cd3ec0822f4a784bc8c8aa4a4affe67d3a6b72a4
8.6 MB 7f5d9abfe429f1ae063259198d665690e655b1459eff3c3bfaf19a8fe62d12ae
2.11.1 (Jul 2019) edgetpu_api_2.11.1.tar.gz
5.1 MB 6067fcc921423c3dc23e1bc12186e42ff954eada50b44826b2167700b50dc69e
1.9.2 (Apr 2019) edgetpu_api_1.9.2.tar.gz
8.3 MB 82324ac028fda707926b756cc97430c8d7f5ffa39940ba74d373de7d9dc90c86
Beta (Mar 2019) edgetpu_api.tar.gz
31.8 MB 52e29f89481e935a9ce2beb0bdafc0495a60f74ef89aa220ddd1142e27adb23e

Source code link

To enable collaboration and advancement in edge ML technology, and allow for expanded platform compatibility, we're proud to share the source code for the following Coral software components:

You might also be interested in the TensorFlow source code.