Raspberry Pi AI HAT+ vs AI Kit vs AI HAT+ 2 (2026): Which Hailo Board to Buy

Quick Answer: For most Raspberry Pi 5 vision projects, buy the AI HAT+ 26 TOPS (Hailo-8) — it’s the fastest Pi accelerator for object detection and the one Frigate users should pick. The AI HAT+ 13 TOPS is the budget option and the direct replacement for the retired AI Kit (same Hailo-8L chip, same speed). The new AI HAT+ 2 (Hailo-10H, 40 TOPS INT4, 8 GB RAM) is the only one that can run an LLM — but independent tests found it slower than the Pi 5’s own CPU for chat, so only buy it if you specifically need vision plus a small language model at the lowest power.

The four boards at a glance

Launch prices from Raspberry Pi and Tom’s Hardware; AI HAT+ 2 reseller price via ComputingForGeeks (checked October 12, 2026). Raspberry Pi raised board prices several times in 2026 because of memory costs, so check the live listing.

Two notes on that table. First, the 40 TOPS of the AI HAT+ 2 is an INT4 figure while the others are INT8 — the numbers are not directly comparable. Second, “own RAM” is the whole story for LLMs: the Hailo-8 and Hailo-8L stream weights over PCIe and have no memory to hold a language model, which is why only the Hailo-10H board can run one at all.

AI Kit vs AI HAT+ 13 TOPS: the same chip in a better package

The AI Kit launched in mid-2024 as a Hailo-8L M.2 2242 module pre-fitted to Raspberry Pi’s M.2 HAT+. A few months later the AI HAT+ put the same chip directly on the HAT board. Performance is identical — Tom’s Hardware notes the 13 TOPS AI HAT+ matches the AI Kit — and the HAT+ version is simpler and slimmer. SparkFun has retired the AI Kit from its catalog and Waveshare lists it as discontinued, both pointing buyers to the AI HAT+.

If you already own an AI Kit, keep it: the software (Hailo runtime, rpicam-apps integration, Frigate’s hailo8l detector) is the same. If you’re buying today, there’s no reason to hunt down leftover AI Kit stock.

Ordering a HAT, a camera and the 27 W power supply separately is where Pi builds lose a week — a free 30-day Amazon Prime trial gets every part there in two days.

AI HAT+ 13 vs 26 TOPS: when the Hailo-8 is worth $40 more

The Hailo-8 is roughly 1.5× faster per frame and has headroom for bigger models and more streams. For a single doorbell or a bird-feeder camera, the 13 TOPS board is enough. For a Frigate NVR with three or more cameras, or if you want to experiment with larger YOLO variants, the 26 TOPS board is the better long-term buy. (Frigate users weighing it against a Coral should read Hailo-8 vs Google Coral.)

AI HAT+ 2: the LLM board that’s slower than the CPU

The AI HAT+ 2 launched in January 2026 with Hailo’s new Hailo-10H chip and 8 GB of dedicated LPDDR4X, so it can hold a small generative model. Raspberry Pi supports models such as DeepSeek-R1-Distill 1.5B, Qwen2/Qwen2.5 1.5B and Llama 3.2 1B through a “hailo-ollama” server.

Independent tests were sobering:

Jeff Geerling’s measurements on the same Pi 5 8 GB, compiled by ComputingForGeeks (October 2026). Other hands-on tests found the same: about 6.9 tok/s on the HAT versus 12 tok/s on the CPU for a 1.5B model.

Where the AI HAT+ 2 does make sense is power and parallelism: the whole Pi peaked at about 7 W under LLM load with the HAT, versus about 12 W running the model on the CPU, and the CPU stays free for other work. It’s also a capable vision accelerator. Tom’s Hardware called it “an interesting prospect” but “a bit of a gamble”. If you want fast local LLMs on a small board, the Jetson Orin Nano Super runs 3B models at over 40 tokens/s.

What else you need

• Raspberry Pi 5: the AI HAT+ boards need the Pi 5’s PCIe connector — they don’t work on a Pi 4. The 4 GB model is enough for Hailo vision; 8 GB if you also want CPU-side LLM experiments.
• 27 W USB-C power supply and the Active Cooler: sustained AI work heats the SoC.
• A camera: Camera Module 3 or the HQ Camera for rpicam pipelines; any IP cameras for Frigate.
• No NVMe on the same lane: the AI HAT+ uses the Pi 5’s only PCIe lane. Boot from a good microSD or USB SSD, or use a third-party PCIe-switch board (shared bandwidth).

Which one should you buy?

The bottom line: the 26 TOPS AI HAT+ is the default pick, the 13 TOPS version is the budget pick and the AI Kit’s successor, and the AI HAT+ 2 is a specialist board you should only buy for its low-power LLM-plus-vision trick — not for speed.