For camera work on a Raspberry Pi 5, such as object detection, segmentation or pose estimation, the AI HAT+ with the Hailo-8 chip (26 TOPS) is the board to buy. The AI HAT+ 2 with the Hailo-10H adds 8 GB of its own memory so it can run small language and vision-language models, but Raspberry Pi itself says its vision performance is comparable to the 26 TOPS AI HAT+. As of October 19, 2026, the AI HAT+ 2 product page lists it at $200, and the AI HAT+ page starts at $70.
The three boards side by side
- AI HAT+ 13 TOPS: Hailo-8L chip. $70 at launch, still the “from” price on Raspberry Pi’s page. Same chip as the older M.2 AI Kit.
- AI HAT+ 26 TOPS: Hailo-8 chip. $110 at its October 2024 launch. Raspberry Pi says it can run several models at once.
- AI HAT+ 2: Hailo-10H chip, rated 40 TOPS at INT4, plus 8 GB of on-board RAM. It launched on January 15, 2026 at $130, as Jeff Geerling reported; Raspberry Pi’s product page now lists $200.
All three fit only the Raspberry Pi 5, conform to the HAT+ spec and come with a 16 mm stacking header so they can sit above the Active Cooler. All three plug into Raspberry Pi’s camera software stack, so detection or segmentation can run as camera post-processing without extra code.
Prices at resellers vary, so check the “Buy” button on Raspberry Pi’s page, which lists approved resellers for your region.
Why the TOPS numbers don’t line up
The 26 TOPS of the Hailo-8 and the 40 TOPS of the Hailo-10H aren’t measured the same way. Raspberry Pi quotes the Hailo-10H figure at INT4, a lower precision that produces bigger numbers. On the vision models people actually run on these boards, Raspberry Pi says the AI HAT+ 2 performs about the same as the 26 TOPS AI HAT+. Don’t buy the AI HAT+ 2 for faster YOLO.
What the AI HAT+ 2 is for
The Hailo-10H can use the 8 GB of RAM on the HAT itself, so it can hold a small language model or vision-language model without taking memory from the Pi. Hailo provides sample models, and the software for supported generative models is in Hailo’s GitHub repository, according to the product page.
Independent testing is less flattering. Jeff Geerling ran every Hailo language model available at launch and found the Pi 5’s own CPU faster in nearly all cases. The Hailo chip was slightly more efficient, since it’s capped at about 3 W. Running vision and a language model at the same time produced errors in his tests. And 8 GB limits you to small models; a 16 GB Pi 5 running llama.cpp on the CPU can load larger quantized ones, slowly.
The AI HAT+ 2 makes sense if you need low-power, offline vision plus a small model on one board, or you’re developing for Hailo-10H hardware. For a chatbot on a Pi, the CPU and more RAM are the simpler route.
How long each will be made
Raspberry Pi commits to producing the AI HAT+ until at least January 2030 and the AI HAT+ 2 until at least January 2036. That matters if you’re designing a product rather than a weekend project.
Choosing quickly
- Camera project, tight budget, one model at a time: AI HAT+ 13 TOPS.
- Camera project with several models or higher frame rates: AI HAT+ 26 TOPS.
- Small LLM or VLM running on the accelerator, low power, no network: AI HAT+ 2.
- Larger models, robot navigation stacks or CUDA software: a Raspberry Pi is the wrong base. Look at NVIDIA’s Jetson boards; see Jetson Orin Nano vs Orin NX.
For how to split work between CPU, GPU and NPU on a robot, see our edge compute guide.