YOLO26 on a Raspberry Pi 5 with the AI Kit (Hailo 8L): Setup, NCNN Export & Real Frame-Rate Benchmarks

If you want the fastest real-time detection on a roughly $150 build, the Raspberry Pi 5 paired with the official AI Kit is currently one of the most honest value-for-performance configurations in edge AI — but only if you ship the right model format. This guide walks through running Ultralytics YOLO26 on a Pi 5, exporting it to NCNN (the fastest CPU format) and to a Hailo HEF for the AI Kit’s 13-TOPS NPU, and then reports only the frame-rate numbers that Ultralytics itself has published — no invented benchmarks.

Why NCNN is the fastest YOLO26 export on a Pi 5

The single most important decision in an edge deployment is not the model — it is the export format. Ultralytics’ official Raspberry Pi guide states plainly that of all supported export formats, NCNN delivers the best inference performance on Raspberry Pi devices because it is highly optimized for ARM-based mobile and embedded platforms. On a Pi 5, the difference is not marginal: the same YOLO26n model runs in roughly a third of the time in NCNN versus ONNX, and in under a quarter of the time versus plain PyTorch. If you install nothing else from this guide, install NCNN.

Setting up YOLO26 on a Raspberry Pi 5

Start with Raspberry Pi OS Bookworm (Debian 12), 64-bit — the Pi 5’s Cortex-A76 processor is aarch64, and only a 64-bit OS is supported. The board’s quad-core CPU runs at 2.4 GHz, and for sustained 24×7 workloads Ultralytics recommends booting from an SSD over the M.2/PCIe slot rather than an SD card, which wears out under continuous writes.

Install the Ultralytics package on the Pi:

sudo apt update
sudo apt install python3-pip -y
pip install -U pip
pip install ultralytics
sudo reboot

No extra drivers are needed for NCNN — export dependencies install automatically the first time you export a PyTorch model.

Export YOLO26 to NCNN and run inference

Exporting a YOLO26n checkpoint to NCNN is a two-line job in Python:

from ultralytics import YOLO

model = YOLO("yolo26n.pt")
model.export(format="ncnn")  # creates 'yolo26n_ncnn_model'

ncnn_model = YOLO("yolo26n_ncnn_model")
results = ncnn_model("https://ultralytics.com/images/bus.jpg")

For live camera inference, Ultralytics pairs the NCNN model with picamera2 (pre-installed on Raspberry Pi OS), looping capture → inference → annotated-frame display. Note that this CPU path does not use the Hailo NPU.

Export YOLO26 to a Hailo HEF for the AI Kit

The AI Kit bundles a Raspberry Pi M.2 HAT+ with a Hailo-8L neural processing unit rated at 13 TOPS INT8, wired to the Pi 5’s PCIe interface. To use it, you do not run a .pt or NCNN model — you compile a Hailo Executable Format (HEF) file on a Linux x86_64 workstation, then copy it to the Pi and run it through HailoRT.

Ultralytics owns the full format="hailo" pipeline: .pt → ONNX → Hailo parse → INT8 optimization → HEF compile. Install Ultralytics plus the Dataflow Compiler (DFC) wheel matching your hardware — Hailo-8 and Hailo-8L use DFC v3.x — then export:

from ultralytics import YOLO

model = YOLO("yolo26n.pt")
output = model.export(format="hailo", name="hailo8l")
print(output)  # yolo26n_hailo_model/

Two practical caveats from the Ultralytics Hailo docs. First, name="hailo8l" is the default target but must match the actual accelerator. Second, HEF compilation is INT8-only and hardware-specific with a fixed input size — resizing must happen on the host per compiled resolution. YOLO26n’s INT8 HEF retains roughly 93% of its PyTorch mAP50, and because quantization shifts YOLO26 confidences down by about 0.05, lowering the confidence threshold to around 0.20 recovers the lost detections.

Real frame-rate benchmarks (from the Ultralytics guide)

The only FPS figures in this article are the ones Ultralytics’ team ran and published in its Raspberry Pi guide, benchmarked with Ultralytics 8.4.108 on a Pi 5 at FP32 precision, 640×640 input (inference time excludes pre- and post-processing). FPS values shown are derived by the author from Ultralytics’ published per-image inference times.

Source: Ultralytics “Quick Start Guide: Raspberry Pi with Ultralytics YOLO26” benchmark table (Ultralytics 8.4.108).

NCNN is the clear winner on CPU at ≈15 FPS. For context, Ultralytics also reports that YOLO26n is about 15% faster than YOLO11n on the same Pi 5 (6.79 → 7.79 FPS, ONNX), while improving mAP from 39.5 to 40.1 — the reason to jump to YOLO26 for edge deployment. Ultralytics likewise cautions that the Hailo path publishes no Pi 5 FPS figure in its own docs, so no Hailo frame-rate is claimed here; reproduce it on your own board and measure your full pipeline (video decode, resize, drawing) rather than trusting a spec sheet.

Pi 5 CPU-only vs. +Hailo 8L vs. Coral M.2

Prices: Raspberry Pi AI Kit $70 list (raspberrypi.com product brief). Hailo-8L = 13 TOPS. Coral figures are market estimates; YOLO26 frame rates are not covered by the Ultralytics guide.

For the fastest documented real-time detection on a fresh ~$150 build, the practical takeaway is: get the Pi 5, add the AI Kit, keep the NCNN path for CPU fallback, and compile a Hailo HEF for the NPU — then benchmark your own full camera pipeline, because end-to-end frames-per-second depends on your app, not just the model.

Reviewed and approved by Prof. Ajay S., professor at a State University in Delhi.

Sources: Ultralytics — Raspberry Pi guide, Hailo integration; Roboflow — What Is YOLO26?; Raspberry Pi — AI Kit documentation; Edge Impulse — Raspberry Pi 5.