How to Run Ultralytics YOLO26 on a Raspberry Pi with a Coral Edge TPU#

A Raspberry Pi is a power-efficient, affordable platform for running computer vision at the edge, but on-device inference is slow even with optimized formats like ONNX or OpenVINO. Pairing the Pi with a Coral Edge TPU coprocessor offloads inference to dedicated hardware and dramatically speeds it up. This guide shows you how to install the runtime, export an Ultralytics YOLO26 model to the Edge TPU format, and run accelerated inference.

Why Use a Coral Edge TPU?#

The Coral Edge TPU is a compact device that adds an Edge TPU coprocessor to your system, enabling low-power, high-performance ML inference for TensorFlow Lite models. It is a great fit for embedded and mobile deployments where a CPU alone cannot keep up:

• Faster inference — the Edge TPU accelerates quantized models far beyond what the Raspberry Pi CPU achieves on its own.
• Low power draw — it delivers high throughput per watt, ideal for battery- or solar-powered deployments.
• Plug-and-play — the USB Accelerator connects over USB 3.0, so no extra hardware integration is required.

Prerequisites#

• Raspberry Pi 4B (2GB or more recommended) or Raspberry Pi 5 (Recommended)
• Raspberry Pi OS Bullseye/Bookworm (64-bit) with desktop (Recommended)
• Coral USB Accelerator
• A non-ARM platform (Google Colab, an x86_64 Linux machine, or the Ultralytics Docker container) for exporting the model, since the Edge TPU compiler is not available on ARM

This guide assumes you already have a working Raspberry Pi OS install with ultralytics and its dependencies installed. If not, follow the quickstart guide first.

With the prerequisites ready, the workflow has three steps: install the Edge TPU runtime on the Pi, export your model on a non-ARM machine, and run inference back on the Pi.

Install the Edge TPU Runtime#

The runtime ships in several builds, so pick the one that matches your operating system. The high-frequency build runs the Edge TPU at a higher clock speed for better performance, but it can cause thermal throttling — use some form of cooling if you choose it.

Download the latest version from here, then install the .deb package:

sudo dpkg -i path/to/package.deb

After installing the runtime, plug your Coral Edge TPU into a USB 3.0 port on the Raspberry Pi so the new udev rule can take effect.

# If you installed the standard version
sudo apt remove libedgetpu1-std

# If you installed the high-frequency version
sudo apt remove libedgetpu1-max

Export Your Model to Edge TPU Format#

To use the Edge TPU, convert your model to a compatible format. Run the export on a non-ARM platform — Google Colab, an x86_64 Linux machine, the official Ultralytics Docker container, or Ultralytics Platform — since the Edge TPU compiler is not available on ARM. See the Export mode for the available arguments.

from ultralytics import YOLO

# Load a model
model = YOLO("path/to/model.pt")  # Load an official model or custom model

# Export the model
model.export(format="edgetpu")

The exported model is saved in the <model_name>_saved_model/ folder as <model_name>_full_integer_quant_edgetpu.tflite.

Run Inference on the Edge TPU#

Before running the model, install the correct libraries on the Raspberry Pi. If TensorFlow is already installed, uninstall it first:

pip uninstall tensorflow tensorflow-aarch64

Then install or update tflite-runtime:

pip install -U tflite-runtime

Now you can run inference:

from ultralytics import YOLO

# Load a model
model = YOLO("path/to/<model_name>_full_integer_quant_edgetpu.tflite")  # Load an official model or custom model

# Run Prediction
model.predict("path/to/source.png")

Find full prediction-mode details on the Predict page.

from ultralytics import YOLO

# Load a model
model = YOLO("path/to/<model_name>_full_integer_quant_edgetpu.tflite")  # Load an official model or custom model

# Run Prediction
model.predict("path/to/source.png")  # Inference defaults to the first TPU

model.predict("path/to/source.png", device="tpu:0")  # Select the first TPU

model.predict("path/to/source.png", device="tpu:1")  # Select the second TPU

Benchmarks#

The figures below were measured with Raspberry Pi OS Bookworm 64-bit and a USB Coral Edge TPU. They show inference time only (pre-/postprocessing excluded) and serve as a relative reference for the acceleration the Edge TPU provides across Pi models and modes.

On average:

• The Raspberry Pi 5 is 22% faster with the standard mode than the Raspberry Pi 4B.
• The Raspberry Pi 5 is 30.2% faster with the high-frequency mode than the Raspberry Pi 4B.
• The high-frequency mode is 28.4% faster than the standard mode.

Conclusion#

A Coral Edge TPU turns a Raspberry Pi into a capable, low-power inference device for Ultralytics YOLO26. Export your model on a non-ARM machine, keep the _edgetpu.tflite suffix, and run it with tflite-runtime on the Pi to get accelerated edge inference. For more deployment options, see the Raspberry Pi guide.

FAQ#