The Raspberry Pi AI HAT+ 2 is an add-on board based on the 40 TOPS Hailo-10H AI accelerator with 8GB of dedicated on-board RAM that brings generative AI capability to Raspberry Pi 5.
While it delivers similar computer vision performance as the first-generation Hailo-8-based Raspberry Pi AI HAT+, the AI HAT+ 2 also adds support for large language models (LLMs) and vision-language models (VLMs) running locally without the need for Internet access. Target applications include offline process control, secure data analysis, facilities management, and robotics.
Raspberry Pi AI HAT+ 2 specifications:
- AI accelerator – Hailo Hailo-10H
- AI accelerator delivering 40 TOPS (INT4) inferencing performance
- Performance for computer vision models comparable to the Raspberry Pi AI HAT+ (26 TOPS)
- 8GB on-board RAM
- Host interface
- PCIe Gen3 x1 FPC connector to Raspberry Pi 5
- 40-pin GPIO header (no signal used by the Hailo-10H, it only extends the GPIO header on the Pi)
- Misc
- Ships with 16mm stacking headers, spacers, and screwsfor installation with the Raspberry Pi 5 active cooler in place
- Optional heatsink
- Dimensions – 64.1 x 56.7 x 5.5mm (Raspberry Pi HAT+ compatible)
- Temperature Range – 0°C to 50°C
- Life cycle – In production until at least January 2036
You’ll need an up-to-date Raspberry Pi OS image to get started, after which the system will automatically detect the Hailo-10H accelerator, which is fully integrated into Raspberry Pi’s camera software stack, notably the rpicam‑apps camera applications, libcamera, and Picamera2. You can find Generative AI models on the Hailo website and on GitHub, where you’ll also find Hailo-Ollama, an Ollama-compatible API written in C++ on top of HailoRT. The best way to get started is probably to check out the relevant documentation on the Raspberry Pi website.
Example of supported models:
- DeepSeek-R1-Distill – 1.5 billion parameters
- Llama3.2 – 1 billion parameters
- Qwen2.5-Coder – 1.5 billion parameters
- Qwen2.5-Instruct – 1.5 billion parameters
- Qwen2 – 1.5 billion parameters
The model won’t quite fill the 8GB of RAM from the module, but we’re told larger models are being readied. Here’s one of the LLM demos.
Output from vision language demo:
Raspberry Pi sent us a sample via DHL last week, but it only arrived to destination a couple of minutes ago as I was about to complete this article. I’ll probably test it this weekend. It is sold for $130 from your favorite Raspberry Pi Reseller, excluding shipping and taxes. It should be a competitor to the Rockchip RK1820/RK1828 LLM/VLM accelerators for which we don’t have clear pricing information.
Jean-Luc started CNX Software in 2010 as a part-time endeavor, before quitting his job as a software engineering manager, and starting to write daily news, and reviews full time later in 2011.
Support CNX Software! Donate via cryptocurrencies, become a Patron on Patreon, or purchase goods on Amazon or Aliexpress. We also use affiliate links in articles to earn commissions if you make a purchase after clicking on those links.
