Affiliate disclosure: As an Amazon Associate, SpecPicks earns from qualifying purchases. Product links in this article are affiliate links; prices and stock shown may vary — see current price on Amazon before ordering.
Yes, if you want small language models and real-time vision running locally on a Pi 5. The AI HAT+ 2 is built on the Hailo-10H accelerator, rated at 40 TOPS (INT4), and carries 8 GB of its own on-board RAM, per Raspberry Pi's launch announcement. That RAM is what the original Hailo-8 AI HAT+ lacked, and it is why this board can run 1–1.5B-parameter LLMs. For vision-only projects Raspberry Pi says performance is "broadly equivalent" to the 26 TOPS AI HAT+, so the upgrade pays off only if you want generative workloads.
Correction (2026-10-17): an earlier version of this article said the AI HAT+ 2 delivers 4 TOPS at 1.2 W, called it a Hailo-8-class board and said it could not run LLMs. All three claims were wrong. It uses the Hailo-10H at 40 TOPS (INT4), and Raspberry Pi markets it for LLMs and vision-language models. Raspberry Pi publishes no power figure for the board, so we have removed ours.
The edge-AI-on-Pi audience and what the HAT+ 2 changes
The Pi has been a favorite edge platform for a decade because it's cheap, sips power and has a mature software stack. What it never had was inference throughput: running a modern vision model on the Pi's CPU means single-digit frames per second and a hot chip. The 2024 PCIe-based AI Kit and AI HAT+ (Hailo-8L at 13 TOPS, Hailo-8 at 26 TOPS) fixed vision. Those boards have no memory of their own, though, so a language model never fit.
The HAT+ 2 changes the accelerator rather than iterating on it. According to Raspberry Pi, it pairs the Hailo-10H with 8 GB of dedicated RAM and "brings generative AI capability to Raspberry Pi 5". At launch Raspberry Pi listed DeepSeek-R1-Distill 1.5B, Llama 3.2 1B, Qwen2.5-Coder 1.5B, Qwen2.5-Instruct 1.5B and Qwen2 1.5B as supported models, with "more (and larger) models" planned.
Audience: makers who want an offline voice assistant, chat or coding helper on a Pi 5; camera and robotics builders who also want on-device scene description; and hobbyists who want local inference without a Jetson's cost or power budget.
Key takeaways
- Hailo-10H, 40 TOPS (INT4), 8 GB on-board RAM, per Raspberry Pi's product page and launch post.
- Runs small LLMs (1–1.5B parameters at launch) through the
hailo-ollamabackend, which keeps the Pi 5's own RAM and CPU free. - Vision performance is "broadly equivalent" to the 26 TOPS AI HAT+, per Raspberry Pi. If you only need detection or pose, the original HAT+ is enough.
- Pi 5 only. A Pi 4 8GB can use USB accelerators but not the HAT+ 2.
- Pair it with fast storage such as the WD_BLACK SN770 NVMe on an NVMe HAT to shorten model-load times.
- Raspberry Pi lists no power figure for the HAT+ 2. Size the supply for the whole system and use the official 27 W USB-C supply.
What is the AI HAT+ 2 and how does it attach?
The AI HAT+ family connects to the Pi 5's PCIe FFC connector. Earlier Pis exposed only USB, GPIO and SPI, none of which can feed a modern accelerator at high frame rates. Raspberry Pi notes that the HAT+ 2 fits on a Pi 5 with the Active Cooler installed.
| Spec | AI HAT+ 2 | AI HAT+ (26 TOPS) |
|---|---|---|
| Base board | Raspberry Pi 5 | Raspberry Pi 5 |
| Accelerator | Hailo-10H | Hailo-8 |
| Rated throughput | 40 TOPS (INT4) | 26 TOPS (INT8) |
| On-board RAM | 8 GB | None |
| Generative AI (LLM/VLM) | Yes, small models | No |
| Launch price | $130 (Jan 2026) | $110 |
| Source | Raspberry Pi launch post | Raspberry Pi AI HAT+ page |
The two TOPS figures use different precisions (INT4 against INT8), so 40 against 26 is not a like-for-like throughput comparison. Raspberry Pi's own statement that vision performance is broadly equivalent is the more useful guide. Its current product page also lists a higher price than the $130 launch price, so check before you buy.
What does it actually run?
The table separates what Raspberry Pi states from what you should expect. We have not benchmarked this board, so it gives no frame-rate or tokens-per-second figures.
| Workload | Pi 5 CPU only | Pi 5 + AI HAT+ 2 |
|---|---|---|
| Object detection / pose (camera stream) | Low frames per second | Real-time, comparable to the 26 TOPS AI HAT+ (per Raspberry Pi) |
| 1–1.5B LLM chat (Llama 3.2 1B, Qwen2.5 1.5B) | Runs, but takes the CPU and system RAM | Supported at launch via hailo-ollama |
| Vision-language / scene description | Impractical | Named as a target workload by Raspberry Pi |
| 7B+ LLMs | Very slow | Not listed at launch; larger models are planned |
| Simple classification | Fine on CPU | Frees the CPU for other tasks |
Models still have to be compiled for Hailo hardware. You get the curated model set and whatever you compile yourself with the Hailo Dataflow Compiler, not arbitrary Hugging Face checkpoints.
Which base board and storage does it need?
- Base board: the Pi 5 is the only supported board. Per Raspberry Pi's Pi 4 product page, the Pi 4 lacks the exposed PCIe lane that the HAT+ family uses. Pi 4 users are limited to USB accelerators such as the Coral USB stick. If you're on a Raspberry Pi 4 8GB, plan to upgrade the base board.
- Storage: LLM weights load from disk, so storage speed matters more here than on a vision-only build. An NVMe HAT with a WD_BLACK 250GB SN770 NVMe avoids SD-card wear and shortens model loads. A SATA Crucial BX500 1TB on a USB adapter works for archive storage.
- Peripheral: a small maker kit such as the Raspberry Pi Zero W Basic Starter Kit covers ancillary hardware for camera-attached secondary nodes.
Power and thermals
Neither Raspberry Pi nor Hailo publishes a board-level power figure for the AI HAT+ 2, so we don't quote one. Size your supply for the whole system: a Pi 5 with the HAT, NVMe storage and a camera under sustained inference needs the official 27 W USB-C supply. Under-voltage shows up as unstable inference that looks like a software bug.
Raspberry Pi designed the board to fit alongside the Pi 5 Active Cooler. Use it, plus a ventilated case, for sustained LLM or video workloads.
How it compares with a Coral TPU or Jetson
| Platform | Rated throughput | Generative AI | Best fit |
|---|---|---|---|
| Raspberry Pi 5 + AI HAT+ 2 | 40 TOPS (INT4), 8 GB on-board RAM | Small LLMs and VLMs | Local chat or voice plus vision on a Pi 5 |
| Raspberry Pi 5 + AI HAT+ (26 TOPS) | 26 TOPS (INT8) | No | Vision-only Pi 5 projects |
| Coral USB Accelerator | 4 TOPS (INT8), per Coral | No | TFLite vision on any Pi, including a Pi 4 |
| NVIDIA Jetson Orin Nano | Higher, with a CUDA GPU | Yes, broader model support | Larger models, CUDA/PyTorch workflows |
The 4 TOPS figure in the old version of this article belongs to the Coral USB Accelerator, not the HAT+ 2.
Verdict matrix
- Buy the HAT+ 2 if… you're on a Pi 5 and want an offline LLM assistant, voice pipeline or scene description, with vision on the same board.
- Buy the 26 TOPS AI HAT+ instead if… your project is vision only. Raspberry Pi says vision performance is broadly the same, and it costs less.
- Use CPU only if… your workload is low frequency, such as a frame every few seconds.
- Step up to a Jetson if… you need 7B-class models today, CUDA, or the broader NVIDIA software ecosystem.
Software toolchain: what you actually work with
On the LLM side, Raspberry Pi's examples use the hailo-ollama backend with an Open WebUI front end. Camera work uses the same stack as the original AI HAT+: libcamera, rpicam-apps and Picamera2. Custom or fine-tuned models go through the Hailo Dataflow Compiler, which turns an ONNX or TensorFlow graph into Hailo's .hef format. Hailo's hailo-apps GitHub repo carries examples and demos.
Common pipeline patterns:
- PiCamera 2 → Hailo → GStreamer output. A straightforward vision pipeline for camera streams.
- Microphone → speech-to-text → small LLM. An offline voice assistant. Raspberry Pi lists speech-to-text and translation as example applications.
- Custom Python inference loop. For prototyping or non-camera inputs.
Comparing the HAT+ 2 to Coral in more detail
Coral's Edge TPU is the closest Pi-compatible alternative for older boards. It attaches over USB (the Coral USB Accelerator) or as an M.2 card, and runs quantized TFLite models natively.
- Cross-Pi compatibility: Coral works on a Pi 4 and older boards. The HAT+ 2 is Pi 5 only.
- Generative AI: Coral has no on-board model memory and isn't aimed at LLMs. The HAT+ 2 is.
- Toolchain: TFLite is broadly supported. Hailo's compiler is more specialized but has first-party vision and LLM model sets.
- Integration: the HAT+ 2 sits on the Pi 5's PCIe lane. A Coral USB dongle sticks out of a USB port and is easy to knock loose.
Common pitfalls with the HAT+ 2
- Trying to run it on a Pi 4. It isn't supported, so buy a Pi 5.
- Undersized power supply. Use the official 27 W USB-C supply.
- Expecting desktop-class LLMs. The launch models are 1–1.5B parameters, so treat it as a local assistant, not a ChatGPT replacement.
- Buying it for vision only. The 26 TOPS AI HAT+ does vision about as well for less.
- Wrong model format. Models must be compiled for Hailo, and a raw ONNX file won't run.
Related guides
- Raspberry Pi AI HAT+ 26 TOPS: specs, setup and verdict
- N100 mini PC for Home Assistant
- Best budget local LLM box: RTX 3060 12GB + Ryzen 5 5600G
- Best SSD for a Steam Deck or emulation handheld
Citations and sources
This piece is an editorial synthesis of Raspberry Pi and Hailo's published specifications. SpecPicks has not benchmarked the board.
— Mike Perry
