Raspberry Pi AI HAT+

schematics

partial (HAT+ mechanical spec public)

accessory-host

documented(Raspberry Pi 5, which scores 3/documented)

toolchain

partial (open Pi driver integration, Hailo compiler gated)

datasheets

brief (public brief)

blobs

required (Hailo firmware required)

retail

open_market (mass-market)

An officially supported open Raspberry Pi software stack and a public product page, on sale worldwide, but the underlying Hailo model-compiler toolchain is registration-gated. The HAT is scored as an accessory: an add-on tracks the platform it completes, and what a builder gets from this HAT is the openness of the Raspberry Pi 5 it plugs into. Its own design files and its own toolchain are each only partly open, and neither describes the system anyone actually uses, so the host's openness is what decides the score. The Pi 5 sits at 3, documented, because it publishes no schematic, only a mechanical drawing, and an accessory cannot offer more openness than the system it completes, so this HAT sits at 3 as well. raspberrypi.com refuses every automated request, so the evidence here is Raspberry Pi's own AI HAT+ product brief, RP-008133-DS, on datasheets.raspberrypi.com, which carries four of the six components on its own. The documentation counts as a brief rather than a full datasheet because that is what it is - six pages of overview, a specification table, a dimensioned drawing and safety text, against the 93-page register-level datasheet the Pi 5 gets. The design files count as partial on the HAT+ Specification, which describes itself as "A guide to designing Hardware-Attached-on-Top of a Raspberry Pi" and licenses its contents for use "solely in conjunction with the Raspberry Pi products" - an interface standard for building your own board, not this board's design. The host pairing comes from the brief tying the HAT to one host and no other: it "communicates using Raspberry Pi 5's PCIe Gen 3 interface" and ships a stacking kit "to enable fitting on Raspberry Pi 5". The toolchain and firmware positions come from Hailo's own repositories rather than the vendor page, HailoRT carrying the open half and the Model Zoo naming the customer-gated compiler. Retail availability comes from The Pi Hut stocking both variants.

• https://datasheets.raspberrypi.com/ai-hat-plus/raspberry-pi-ai-hat-plus-product-brief.pdf recorded 2026-08-14

  Raspberry Pi's own AI HAT+ product brief, RP-008133-DS, published October 2024 - six pages, of which one is the specification table and one a dimensioned drawing marked "All dimensions are approximate and for reference purposes only ... should not be used for producing production data". Its own text ties the HAT to a single host - "an add-on board with a built-in Hailo AI accelerator for Raspberry Pi 5", "communicates using Raspberry Pi 5's PCIe Gen 3 interface", "Supplied with 16mm stacking header, spacers, and screws to enable fitting on Raspberry Pi 5 with Raspberry Pi Active Cooler in place" - and records "Conforms to Raspberry Pi HAT+ specification" plus "Fully integrated into Raspberry Pi's camera software stack"

• https://datasheets.raspberrypi.com/hat/hat-plus-specification.pdf recorded 2026-08-14

  The Raspberry Pi HAT+ Specification, RP-008281-DS, 21 pages dated 2024-12-05 - "A guide to designing Hardware-Attached-on-Top of a Raspberry Pi", covering power states, the ID EEPROM, HAT+ classes and stackability. CC BY-ND, and the legal notice grants use "solely in conjunction with the Raspberry Pi products. All other use of the RESOURCES is prohibited". A mechanical and electrical interface standard for building your own add-on, which is the rubric's `partial`, and not the design of the board being sold

• https://datasheets.raspberrypi.com/rpi5/raspberry-pi-5-mechanical-drawing.pdf recorded 2026-08-14

  RP-008347-DS, the only board-level design document Raspberry Pi publishes for the Pi 5, and the evidence that puts the host at 3/documented rather than 4/open_toolchain. The reduced schematics path under the same rpi5 prefix 404s while the rpi4 one answers 200

• https://raw.githubusercontent.com/hailo-ai/hailort/master/README.md recorded 2026-08-14

  The open half of the toolchain and the source of the firmware requirement - "libhailort, pyhailort & hailortcli - distributed under the MIT license" with hailonet under LGPL 2.1, while the HailoRT PCIe driver "includes the Hailo-8 firmware that runs on the Hailo device and manages its boot and control"

• https://raw.githubusercontent.com/hailo-ai/hailo_model_zoo/master/README.rst recorded 2026-08-14

  The gated half - "Install Hailo Dataflow Compiler and enter the virtualenv. In case you are not Hailo customer please contact hailo.ai". The model compiler this accelerator needs is customer-gated even though the Model Zoo itself "is released under the MIT license"

• https://thepihut.com/products/raspberry-pi-ai-hat-plus recorded 2026-08-14

  Both variants in stock at a mass-market reseller and addable to a cart with no account or approval - 13 TOPS at GBP67.20 and 26 TOPS at GBP105.60 inc VAT

• https://www.raspberrypi.com/products/ai-hat/ recorded 2026-06-22

  Official product page: Hailo-8 (26 TOPS) / Hailo-8L (13 TOPS), HAT+ form factor, from $70. Not re-fetchable - raspberrypi.com answers HTTP 403 to every automated request, so the source keeps its recorded read

First-party Raspberry Pi accessory sold through the global Pi distributor network with strong community/tutorial coverage. The product page reads "Available now from $70", carries Buy links against the 13 TOPS and 26 TOPS variants and an Approved Resellers list, and commits to production until at least January 2030.

• https://www.raspberrypi.com/products/ai-hat/ recorded 2026-08-13

  "Available now from $70"; Buy links for the 13 TOPS and 26 TOPS variants; "Our Approved Resellers"; "Raspberry Pi AI HAT+ will remain in production until at least January 2030"

Mid-tier vision-focused NPU; strong for CNN inference but not large LLMs, and dependent on the host Pi for memory. The page offers "Hailo-8 or Hailo-8L accelerator offering 26 TOPS or 13 TOPS inferencing performance respectively" and describes the workload as camera post-processing (object detection, image segmentation, pose estimation) rather than generative models. That places it above raspberry-pi-5 and the 6 TOPS RK3588 boards on throughput and below the Jetson Orin parts on what it can actually run.

• https://www.raspberrypi.com/products/ai-hat/ recorded 2026-08-13

  "Hailo-8 or Hailo-8L accelerator offering 26 TOPS or 13 TOPS inferencing performance respectively"; "fully integrated into Raspberry Pi's camera software stack ... object detection, image segmentation, and pose estimation"