Raspberry Pi AI HAT+ 2
A PCIe AI accelerator for the Raspberry Pi 5, using the Hailo-10H chip to deliver 40 TOPS of on device AI inference with 8GB of dedicated memory.
The Raspberry Pi AI HAT+ 2 is a PCIe add-on board for the Raspberry Pi 5 that brings dedicated on device AI acceleration. Released in January 2026, it connects through the Pi's PCIe interface for a direct, high bandwidth link to the chip.
The Hailo-10H chip
The HAT+ 2 is built around the Hailo-10H, a chip designed specifically for edge AI inference. It delivers 40 TOPS of INT4 performance, a significant step up from previous edge hardware.
What makes the Hailo-10H different from its predecessor, the Hailo-8 found in the original AI HAT+, is the onboard memory. The Hailo-10H carries 8GB of LPDDR4X memory directly on the module. That matters because AI models, particularly language and vision models, require substantial memory to load and run. The original HAT+ had to share the Pi's RAM with everything else running on the board. The HAT+ 2 does not. It brings its own.
The whole module runs at a maximum of 3 watts. For a chip capable of running small generative AI models entirely on device, that power figure is remarkable.
What this unlocks
The original AI HAT+ and the Google Coral USB Accelerator are built for object detection. The HAT+ 2 targets a different class of workload: small language models and on device generative AI that previously required cloud access or significantly more expensive hardware. Running a local language model on a device that draws under 10 watts and sends nothing to an external server is what this opens up.
The AI HAT+ 2 is in production until at least January 2036, with Raspberry Pi Ltd committed to ongoing availability. For anyone building a setup expected to run for years, that kind of longevity commitment matters.