Skip to content

THE FIELD JOURNAL

Raspberry Pi AI HAT+ 2: Specs, Models, Setup and Real LLM Speeds

By Local AI Scout · Beginner guide

Hand-drawn illustration of a camera image window and a chat window feeding a small AI chip under a laptop screen, for a guide to the Raspberry Pi AI HAT+ 2

What is the Raspberry Pi AI HAT+ 2, and is it worth buying?

The Raspberry Pi AI HAT+ 2 is a $200 add-on board for the Raspberry Pi 5 with a Hailo-10H accelerator rated at 40 TOPS (INT4) and 8GB of its own RAM, so it runs small LLMs, vision-language models and speech models offline. It suits camera and low-power projects, but for chat speed the Pi 5's own CPU was faster.

This guide follows one reader: someone with an 8GB Raspberry Pi 5 running Raspberry Pi OS who wants either a private assistant that answers questions with no internet connection, or a camera project that can describe what it sees, and who is weighing $200 for the HAT against running a model on the Pi's own processor. It is part of our guide to AI on a Raspberry Pi, which maps every route.

Choose the AI HAT+ 2 if:

Skip it and use the Pi 5 CPU if:

AI HAT+ 2 specs and price in October 2026

An NPU, or neural processing unit, is a chip built only for running AI models, the same idea as the NPU in an AI PC. TOPS means trillions of operations per second, and the number depends on precision: INT4 means 4-bit whole numbers, which are faster but coarser than 8-bit (INT8). Hailo rates the chip at 40 TOPS at INT4 and 20 TOPS at INT8.

FeatureRaspberry Pi AI HAT+ 2
AcceleratorHailo-10H NPU
AI performance40 TOPS (INT4); 20 TOPS (INT8) per Hailo
Memory8GB on the board, separate from the Pi's RAM
Model sizeLLMs and VLMs up to about 6 billion parameters, per Raspberry Pi
ConnectionPi 5 PCIe port by ribbon cable, plus GPIO stacking header; PCIe Gen 3.0 applied automatically
Works withRaspberry Pi 5 only, on 64-bit Raspberry Pi OS Trixie
Chip power2.5W typical (Hailo's figure for the chip)
SizeAbout 66 x 56.5 mm; about 14 mm high with the heatsink's push pins
Operating temperature0°C to 50°C ambient
In the boxOptional heatsink, 16mm stacking header, spacers and screws
Price$130 at launch; $200 listed in October 2026
LaunchedJanuary 15, 2026
In production untilAt least January 2036

The price has moved. Raspberry Pi launched it at $130 on January 15, 2026, and its April 1, 2026 round of memory-driven price increases added $50. Its product page and July 2026 product brief now list $200; I found no announcement for the last $20. Raspberry Pi says its Approved Resellers abide by its published maximum prices.

Does the AI HAT+ 2 work with a Raspberry Pi 4 or other boards?

No. The product brief says it shall only be connected to a Raspberry Pi 5 via the PCIe interface and GPIO header. Raspberry Pi does not publish a power figure for the whole board; Hailo's own write-up lists 2.1W average power for the chip running Qwen2 1.5B.

How the AI HAT+ 2 runs a model

An LLM, or large language model, predicts text one token (a word or word piece) at a time; our explainer on how LLMs work covers the basics. On the HAT+ 2, Hailo says the entire LLM pipeline is offloaded to the accelerator, and the model loads into the HAT's own memory, leaving the host Raspberry Pi 5 free to handle other tasks.

Diagram of the Raspberry Pi AI HAT+ 2: the Pi 5 CPU sends a prompt over the PCIe ribbon cable to the Hailo-10H, which runs the model from its own 8GB while the Pi's RAM stays free
The model lives in the HAT's 8GB, so the Pi's own memory stays free for the operating system, the camera and your app.

Models are squeezed to fit through quantization, storing each weight in fewer bits. Hailo's LLMs use A8W4, 8-bit activations with 4-bit weights, and a 2,048-token context, the text a model can keep in view at once. Hailo says that is about 1,536 words, and a 1.5B model with that context needs about 1.2GB.

Which models does the AI HAT+ 2 run?

Only models Hailo has compiled for the chip. They come as HEF files, Hailo's own executable format, not the GGUF files Ollama and llama.cpp use, and Hailo's server does not let users upload models. Raspberry Pi's launch post listed five LLMs of 1 to 1.5 billion parameters and promised more, and larger, models soon after.

Versions matter here. Raspberry Pi's docs install version 5.1.1 of Hailo's GenAI package, and its software repository ships HailoRT 5.1.1 for the Hailo-10H. Hailo's current Model Zoo GenAI is 5.4.0, and it notes that some precompiled models are compatible only with v5.4.0. The table shows Hailo's list in October 2026, and its last column shows what runs on Raspberry Pi's standard install: the 5.1.1 release serves five LLMs through hailo-ollama and runs Qwen2-VL 2B and Whisper Base through Hailo's own apps, while the other models need a newer HailoRT than Raspberry Pi ships.

Model (parameters)TypeSizeHailo's speed, 5.4.0Runs on the Pi's 5.1.1 install
Qwen2 1.5B InstructChat LLM1.56 GB8.06 tok/sYes, hailo-ollama
Qwen2.5 1.5B InstructChat LLM1.64 GB7.35 tok/sYes, hailo-ollama
Qwen2.5 Coder 1.5B InstructCoding LLM1.64 GB8.13 tok/sYes, hailo-ollama
DeepSeek R1 Distill Qwen 1.5BReasoning LLM2.37 GB7.96 tok/sYes, hailo-ollama
Llama 3.2 1B InstructChat LLM1.79 GB9.89 tok/sNo
Qwen3 1.7B InstructChat LLM1.79 GB4.78 tok/sNo
Qwen2 1.5B Function CallingTool-use LLM2.99 GB6.69 tok/sNo
Llama 3.2 3B InstructChat LLM3.13 GB2.67 tok/s (5.1.1)Yes, hailo-ollama; removed in 5.2.0
Qwen2-VL 2B InstructVision-language2.18 GB7.04 tok/sYes, Hailo apps
Qwen3-VL 2B InstructVision-language2.18 GB4.74 tok/sNo
Whisper Tiny (39M)Speech to text78 MB48.14 tok/sNo
Whisper Base (74M)Speech to text155 MB25.32 tok/sYes, Hailo apps
Whisper Small (244M)Speech to text388 MB10.61 tok/sNo

These speeds are Hailo's own figures, in tokens per second of output, and the list does not state the host system or prompt length. Its blog lists Qwen2 1.5B at 9.45 tokens per second, with time to first token measured on a 96-token prompt. Time to first token, the wait before text starts, is 0.32 to 0.73 seconds for the LLMs. The changelog also shows Stable Diffusion 1.5 was dropped in 5.1.1, so Hailo's current list has no image generator.

Bar chart of Hailo's published speeds for the Hailo-10H chip on the Raspberry Pi AI HAT+ 2, from 9.89 tokens per second for Llama 3.2 1B to 4.78 for Qwen3 1.7B
On Hailo's own 5.4.0 numbers every language model writes at under 10 tokens per second; the two starred models need a newer HailoRT than Raspberry Pi ships.

Can the AI HAT+ 2 run vision-language models?

Yes. A VLM, or vision-language model, takes an image plus a question and answers in text. Qwen2-VL 2B reads a 336 x 336 pixel image, and Hailo lists a 1.02-second time to first token for it. Qwen3-VL 2B supports single-frame image input only, not video. VLMs do not run through Hailo's Ollama-style server, which is limited to LLMs; Raspberry Pi points to Hailo's hailo-apps repository instead.

Its VLM Chat app shows a live camera feed; you press Enter to freeze a frame and type a question about it, and the app uses Qwen2-VL-2B-Instruct on the Hailo-10H. These commands come from Hailo's installation guide and the app's README, run after the driver setup below:

git clone https://github.com/hailo-ai/hailo-apps.git
cd hailo-apps
./install.sh
source setup_env.sh
python -m pip install -e ".[gen-ai]"
python -m hailo_apps.python.gen_ai_apps.vlm_chat.vlm_chat --input rpi

Does it handle speech, and do the vision models still work?

Speech to text, yes: Whisper Tiny, Base and Small are on Hailo's list, and Whisper Base is the one in the 5.1.1 release. Hailo's Voice-to-Action demo runs Whisper-Base on the Hailo chip and Piper text to speech on the CPU, and was tested with the Raspberry Pi apt packages. Vision models such as YOLO object detection also work, as Raspberry Pi's rpicam-apps instructions cover the AI HAT+ 2. Not every demo file has a Hailo-10H model yet: the YOLOv8 file loads a YOLOv8m model on the Hailo-10H, while the YOLOv6 file lists only Hailo-8 and Hailo-8L models, so start with YOLOv8. Hailo's guide supports a vision pipeline and a GenAI app running together, but not two GenAI apps.

How to set up the AI HAT+ 2 and run your first LLM

These steps follow Raspberry Pi's AI HAT guide and AI software page. You need 64-bit Raspberry Pi OS Trixie, an internet connection for downloads, and spare storage for models.

  1. Fit the cooler, heatsink and HAT. With the Pi unplugged, fit the Active Cooler and press the HAT's heatsink on; Raspberry Pi recommends both, the heatsink to stop the HAT slowing down from overheating. Screw on the spacers, press on the stacking header, plug the ribbon cable into the Pi with its metal contacts facing the USB ports, then screw the HAT down and clip the cable into it.
  2. Update the system and firmware. Run the first block of commands below, which includes rpi-eeprom-update for the latest firmware, then reboot.
  3. Install the Hailo-10H package. Install dkms and hailo-h10-all, then reboot. The AI HAT+ and AI Kit use hailo-all instead, and the two packages can't co-exist.
  4. Check the HAT is detected. hailortcli fw-control identify should print a firmware version and device architecture. Serial number, part number and product name may show N/A, which the docs call expected.
sudo apt update
sudo apt full-upgrade -y
sudo rpi-eeprom-update -a
sudo reboot

sudo apt install dkms
sudo apt install hailo-h10-all
sudo reboot

hailortcli fw-control identify
dmesg | grep -i hailo

Run your first model with hailo-ollama

hailo-ollama is Hailo's model server: it downloads models from Hailo and runs them on the HAT. With the drivers working, three steps get you a reply.

  1. Install the hailo-ollama server. Download the 5.1.1 Hailo Model Zoo GenAI package that Raspberry Pi's docs link, and install it with dpkg.
  2. Start the server and pull a model. Run hailo-ollama in one terminal. In a second, list the available models and download one; the docs use qwen2:1.5b as the example.
  3. Ask it something. Send a chat request with curl. The first answer takes a few seconds while the model loads; Hailo lists a 7.37-second load time for Qwen2 1.5B in 5.1.1.
sudo dpkg -i hailo_gen_ai_model_zoo_5.1.1_arm64.deb
hailo-ollama

# in a second terminal
curl --silent http://localhost:8000/hailo/v1/list

curl --silent http://localhost:8000/api/pull \
     -H 'Content-Type: application/json' \
     -d '{ "model": "qwen2:1.5b", "stream" : true }'

curl --silent http://localhost:8000/api/chat \
     -H 'Content-Type: application/json' \
     -d '{"model": "qwen2:1.5b", "messages": [{"role": "user", "content": "Translate to French: The cat is on the table."}]}'
Info card of Raspberry Pi AI HAT+ 2 setup commands: install hailo-h10-all, check the board with hailortcli, install the 5.1.1 model zoo package and start hailo-ollama
Four commands get you from a fitted HAT to a running model server; the most common slip is installing hailo-all, which is for the older boards.

Model names change between releases: Hailo renamed qwen2.5-instruct:1.5b to qwen2.5:1.5b in 5.2.0, so copy names from the list command rather than from a tutorial.

Add a chat window with Open WebUI

Open WebUI is a browser chat interface; our Open WebUI setup guide explains it. On the Pi it must run in Docker, a tool for running apps in sealed containers, because Open WebUI is incompatible with Python 3.13, which Trixie uses. After installing Docker as the docs describe, start it with hailo-ollama already running, then open http://127.0.0.1:8080:

docker pull ghcr.io/open-webui/open-webui:main
docker run -d -e OLLAMA_BASE_URL=http://127.0.0.1:8000 -v open-webui:/app/backend/data --name open-webui --network=host --restart always ghcr.io/open-webui/open-webui:main
docker logs open-webui -f

Does the AI HAT+ 2 have an Ollama or OpenAI-compatible API?

Partly. An API is the address other programs send requests to. hailo-ollama is an Ollama-compatible API written in C++, on port 8000 rather than Ollama's 11434, with endpoints such as /api/chat and /api/generate. Hailo's blog calls the runtime's REST API compatible with Ollama and OpenAI, and the server's source has an OpenAI-style /v1/chat/completions route, but the usage guide documents only the Ollama-style endpoints.

Its usage guide lists the limits: a single simultaneous connection with one pending request, no LoRA adapters and no Ollama CLI. For a private assistant, note that the 5.1.1 server listens on 0.0.0.0, port 8000, meaning every network interface, so other devices on your network can connect unless something blocks them. I found no login or API key in Hailo's docs. In 5.1.1 the address is set in a config file, hailo-ollama.json; release 5.3.0 replaced it with an OLLAMA_HOST setting.

Is the AI HAT+ 2 faster than the Pi 5 CPU for LLMs?

No, on the published evidence. The one like-for-like test I know of is Jeff Geerling's, on an 8GB Pi 5 on launch day. He ran the same models on the HAT, using Hailo's model files, and on the CPU, using Ollama's, and measured power at the wall for the whole Pi. Tokens per second is how fast text appears; our local vs cloud speed guide explains what feels usable.

Model (January 2026 test)AI HAT+ 2Pi 5 CPU (8GB)Tokens per watt, HAT vs CPU
Qwen2 1.5B8.25 tok/s12.47 tok/s1.196 vs 1.014
Qwen2.5 Instruct 1.5B6.864 tok/s11.41 tok/s0.967 vs 0.975
Qwen2.5 Coder 1.5B8.127 tok/s9.77 tok/s1.145 vs 0.872
DeepSeek R1 Distill Qwen 1.5B6.851 tok/s10.99 tok/s0.993 vs 0.931
Llama 3.2 3B2.658 tok/s5.55 tok/s0.409 vs 0.5

The CPU won every speed test, and the HAT was slightly more efficient on most models. Power explains it: the whole Pi peaked at 7.1W with the HAT working and 12.3W with the CPU working, and idled at 4.2W in the HAT setup against 3.1W in the CPU setup. His HAT numbers closely match Hailo's own, 8.25 tokens per second against Hailo's 8.18 for Qwen2 1.5B in release 5.1.1, so the chip performs as advertised. The CPU is just quicker at these sizes.

Bar chart comparing Raspberry Pi AI HAT+ 2 and Pi 5 CPU speeds: Qwen2 1.5B 8.25 vs 12.47, Qwen2.5 Coder 8.127 vs 9.77, Llama 3.2 3B 2.658 vs 5.55 tokens per second
On the same models, the Pi 5's CPU wrote faster every time; the HAT's case rests on lower power and a free CPU, not speed.

Two caveats: the test is from January, and Hailo's newer releases list different speeds for some models; and each side used its platform's default model files, not identical ones. Raspberry Pi's own blog introduced an October 2026 partner post from Sixfab and DEEPX with the line that most edge AI/ML workloads belong on the CPU, but accelerators are important for the small minority of workloads that the CPU can't accommodate.

Beyond speed: model size, memory, power and ease

Model size. The HAT is capped by its 8GB and by Hailo's list. The 5.1.1 package Raspberry Pi links tops out at Llama 3.2 3B; in Hailo's current list the largest LLM has 1.7 billion parameters and the vision-language models 2 billion. On the CPU, the limit is the Pi's RAM: Geerling ran Qwen3 30B A3B compressed to fit in about 10GB on a 16GB Pi 5 with llama.cpp.

Memory. On an 8GB Pi 5, a CPU model shares RAM with the system and your apps. The HAT keeps the model off the Pi's RAM entirely, which is its clearest advantage for a camera project that also needs memory.

Power. In Geerling's test the HAT setup peaked about 5W lower than the CPU but idled about 1W higher, so the saving only counts on a device that is busy most of the time.

Ease. The CPU route uses ordinary Ollama with any model that fits in RAM. The HAT route means Hailo's packages and model names, a separate server and a fixed 2,048-token context. For a private chat assistant, the CPU is simpler and faster.

What are the AI HAT+ 2's limits?

Common questions about the Raspberry Pi AI HAT+ 2

How much does the Raspberry Pi AI HAT+ 2 cost?

Raspberry Pi launched the AI HAT+ 2 at $130 on January 15, 2026. Its April 1, 2026 round of memory-driven price increases added $50, and its product page and July 2026 product brief now list $200; I found no announcement for the last $20. Raspberry Pi says its Approved Resellers abide by its published maximum prices.

Does the AI HAT+ 2 work with a Raspberry Pi 4?

No. Raspberry Pi says the AI HAT+ 2 shall only be connected to a Raspberry Pi 5, through its PCIe interface and GPIO header. Its software instructions also assume a Raspberry Pi 5 running the 64-bit version of Raspberry Pi OS Trixie with up-to-date firmware.

Can the AI HAT+ 2 run models from the Ollama library?

No. It runs only models Hailo has compiled for the Hailo-10H, served by Hailo's own hailo-ollama server on port 8000. That server speaks the Ollama API, so tools such as Open WebUI can connect, but you cannot pull ordinary GGUF models or upload your own.

Is the AI HAT+ 2 faster than the Raspberry Pi 5 CPU for LLMs?

Not in the one like-for-like test I know of. Jeff Geerling ran the same models on an 8GB Pi 5 in January 2026, and the CPU was faster every time, for example 12.47 against 8.25 tokens per second on Qwen2 1.5B. The HAT drew less power and left the CPU free.

What is the difference between the AI HAT+ and the AI HAT+ 2?

The AI HAT+ uses a Hailo-8L or Hailo-8 rated at 13 or 26 TOPS (INT8), relies on the Pi's own memory and does not support LLMs or vision-language models. The AI HAT+ 2 has a Hailo-10H rated at 40 TOPS (INT4), or 20 at INT8, with its own 8GB of memory, and runs both. Vision performance is broadly similar.

What is the biggest model the AI HAT+ 2 can run?

Raspberry Pi says its 8GB of memory fits LLMs and vision-language models of up to about 6 billion parameters. In practice you are limited to Hailo's compiled models. The 5.1.1 package that Raspberry Pi links includes Llama 3.2 3B, while Hailo's current list tops out at 2 billion parameters, all with a 2,048-token context.

My verdict for an 8GB Pi 5 owner: buy it for camera projects, not for chat

If you want a private offline assistant, keep your $200. The Pi 5's CPU ran every model in the comparison faster, runs any Ollama model that fits in its RAM, and needs no add-on board. If your project is a camera that spots things and answers questions about them, the AI HAT+ 2 earns its place: the vision model and a 2B vision-language model run on the HAT while the Pi's CPU and 8GB stay free for your code.

Your next step: try a 1.5B model on the CPU first with our Raspberry Pi LLM guide, and if you still want a camera that understands what it sees, compare boards in the AI HAT buying guide. For every other route, from the AI Camera to voice projects, go back to the Raspberry Pi AI hub.

Sources

Keep exploring: Apps · Models · All guides

THE SCOUT LETTER

Good finds.
Less noise.

Useful local AI apps, practical guides and the updates worth your time. A thoughtful note in your inbox.

Free to join. Unsubscribe with one click.