[Hardware] Raspberry Pi AI HAT+ 2: Can a Raspberry Pi 5 Run Local LLMs?

Disclosure: This post contains affiliate links. As an Amazon Associate I earn from qualifying purchases. See our affiliate disclosure.

Quick answer: The Raspberry Pi AI HAT+ 2 can run
small local LLMs (1 to 2 billion parameters) on a Raspberry Pi 5, but
independent reviewers found it generates text more slowly than the Pi
5’s own CPU. Buy it if you need an offline assistant, speech-to-text or
a vision-language model that leaves the Pi’s CPU and RAM free, or runs
on a tight power budget, ideally alongside camera-based vision work. If
you just want the fastest, most flexible local chatbot on a Pi, a 16GB
Raspberry Pi 5 running llama.cpp or Ollama on the CPU is the better buy
for now.

The Raspberry Pi AI HAT+ 2 is Raspberry Pi’s first add-on built for
generative AI. It pairs a Hailo-10H accelerator with 8GB of its own
memory, so models live on the HAT instead of in the Pi’s RAM. The
reality in October 2026 is more mixed than that pitch, and the price has
climbed since launch. Every price, spec and date was checked on
October 23, 2026, and performance numbers are credited
to the people who measured them.

Comparison table

Option	Accelerator	AI memory	LLM / generative AI	Price (checked Oct 23, 2026)	Best for
Raspberry Pi AI HAT+ 2	Hailo-10H, 40 TOPS (INT4)	8GB on the HAT	Yes: 1 to 2B LLMs, VLMs, Whisper	$200	Offloading small GenAI models plus vision, low power
Raspberry Pi AI HAT+ (13 / 26 TOPS)	Hailo-8L / Hailo-8	Uses Pi memory	No (vision models only)	From $70	Object detection, pose, segmentation
Raspberry Pi 5 16GB, CPU only	4x Cortex-A76 at 2.4GHz	16GB shared	Yes, via llama.cpp or Ollama (GGUF)	$305 (board only)	Fastest, most flexible local LLMs on a Pi
Raspberry Pi 5 8GB + AI HAT+ 2	Hailo-10H + CPU	8GB + 8GB	Yes, on either the HAT or the CPU	$175 + $200	Running an LLM while the CPU does other work

Prices are US list prices from raspberrypi.com, except the 8GB Pi 5,
which is the April 2026 list price reported by CNX Software. Local
prices, taxes and reseller stock vary.

Raspberry Pi 5 8GB: Check price on Amazon (paid link)

Specs and key facts

• Accelerator: Hailo-10H neural network accelerator,
  rated at 40 TOPS (INT4) by Raspberry Pi.
• Memory: 8GB of dedicated on-board
  RAM (LPDDR4X, per Jeff Geerling’s review). Models load into
  this memory, not the Pi’s.
• Vision performance: Raspberry Pi says computer
  vision performance is “broadly equivalent” to the 26 TOPS AI HAT+.
• Host: Raspberry Pi 5 only. It connects over the Pi
  5’s PCIe connector, and needs 64-bit Raspberry Pi OS (Trixie).
• Power: Jeff Geerling reports the chip runs at a
  maximum of about 3W.
• In the box: optional heatsink, 16mm stacking
  header, spacers and screws, so it fits over the Raspberry Pi Active
  Cooler.
• Camera integration: works with Raspberry Pi’s
  camera stack (libcamera, rpicam-apps and Picamera2).
• Launch: announced and on sale January 15,
  2026 at $130.
• Price history: Raspberry Pi raised the price by
  $50 on April 1, 2026, blaming memory costs, which took
  it to $180. The official product page now lists $200
  (checked October 23, 2026). We couldn’t find an official announcement
  for the extra $20, so treat $200 as the current list price and expect
  reseller prices to differ.
• Production lifetime: Raspberry Pi says it will stay
  in production until at least January 2036.

Who it’s for (and who should
skip it)

Good fit:

• Makers building always-on, offline devices. Think a
  voice-controlled home assistant, a kiosk that answers questions about a
  fixed document set, or a robot that describes what its camera sees. The
  HAT runs the model while the Pi’s CPU handles GPIO, networking and your
  own code.
• Power-constrained projects. Reviewers measured
  lower power use with the HAT than with CPU inference, which matters on
  battery or solar.
• Projects that mix vision and language. The HAT+ 2
  is the only official Pi add-on that can do object detection and
  run an LLM or VLM, though running both at once was still rough in early
  reviews (see caveats).
• Owners of low-RAM Pi 5 boards. CNX Software ran
  every model on a 2GB Pi 5.

Probably skip it if:

• You want the fastest local chatbot on a Pi. Both
  Jeff Geerling and CNX Software found the Pi 5’s CPU produced tokens
  faster than the HAT in nearly every test.
• You want bigger models. The HAT’s model list tops
  out around 2B parameters today. A 16GB Pi 5 can run much larger
  quantized models on the CPU.
• You only need vision. The original AI HAT+ (from
  $70) or the $70 Raspberry Pi AI Camera does object detection for
  less.
• You need an NVMe SSD on the same Pi. Both the AI
  HATs and the M.2 HAT+ use the Pi 5’s single PCIe connector, so you can’t
  use them at the same time without third-party hardware.

What’s new compared
with the original AI HAT+

The original AI HAT+, launched in late 2024, uses the Hailo-8L (13
TOPS) or Hailo-8 (26 TOPS). It’s great for vision models such as YOLO
object detection and pose estimation, but Raspberry Pi says it “lacks
the capability” to run generative AI models.

The AI HAT+ 2 changes two things:

1. A newer chip. The Hailo-10H architecture is
   designed for transformer-style models such as LLMs and VLMs, not just
   convolutional vision networks.
2. Its own RAM. With 8GB on board, the HAT holds model
   weights itself. On the original HAT+, there was no room for a language
   model at all.

For vision work, Raspberry Pi says the new board performs roughly
like the 26 TOPS HAT+. Existing camera apps should carry over, but a
Raspberry Pi engineer noted in the launch post’s comments that you need
model files compiled for the Hailo-10H. The software packages also
differ: the original HATs use hailo-all, the HAT+ 2 uses
hailo-h10-all, and Raspberry Pi’s docs say the two packages
can’t be installed side by side.

Which models it supports

At launch, Raspberry Pi listed five LLMs: DeepSeek-R1-Distill (1.5B),
Llama 3.2 (1B), Qwen2.5-Coder (1.5B), Qwen2.5-Instruct (1.5B) and Qwen2
(1.5B). Reviewers in January also ran a Llama 3.2 3B build through
hailo-ollama.

Hailo’s GenAI model zoo on GitHub (checked October 23, 2026) now
lists these Hailo-10H models:

• Text LLMs: DeepSeek-R1-Distill-Qwen-1.5B,
  Llama3.2-1B-Instruct, Qwen2-1.5B-Instruct, a function-calling Qwen2 1.5B
  variant, Qwen2.5-1.5B-Instruct, Qwen2.5-Coder-1.5B-Instruct and
  Qwen3-1.7B-Instruct.
• Vision-language models: Qwen2-VL-2B-Instruct and
  Qwen3-VL-2B-Instruct.
• Speech-to-text: Whisper Tiny, Base and Small.

A few limits are worth knowing before you buy:

• Context length is 2,048 tokens for every LLM in
  Hailo’s list. That’s fine for short Q&A and commands, but tight for
  long documents or chat history.
• You can’t load any model you like. Models must be
  compiled into Hailo’s HEF format. Hailo-Ollama only serves models from
  Hailo’s own list, and you can’t upload your own GGUF files.
• Hailo-Ollama is text-only. Hailo’s docs say it
  doesn’t support VLMs, LoRA adapters or the Ollama command line. VLMs and
  Whisper run through HailoRT’s C++ or Python APIs and the examples in
  Hailo’s hailo-apps repository instead. In May 2026, a Hailo staff member
  confirmed on the Hailo Community forum that Hailo-Ollama is designed for
  LLM text generation only.
• Fine-tuning is possible, but not simple. Raspberry
  Pi says LoRA adapters can be compiled with Hailo’s Dataflow Compiler. It
  isn’t a one-click process.

How setup works

Raspberry Pi’s official AI software guide covers the whole process.
In short:

1. Assemble the hardware. Fit the Active Cooler if you
   use one, then mount the HAT on the PCIe connector with the supplied
   stacking header and spacers. Attach a camera first if you plan to do
   vision work.
2. Update the system. Use 64-bit Raspberry Pi OS
   Trixie, then run sudo apt update,
   sudo apt full-upgrade -y and
   sudo rpi-eeprom-update -a, and reboot. Unlike the old AI
   Kit, you don’t need to enable PCIe Gen 3 by hand, because the docs say
   that’s applied automatically for the AI HATs.
3. Install the Hailo stack. Run
   sudo apt install dkms and then
   sudo apt install hailo-h10-all. This installs the kernel
   driver, firmware, HailoRT runtime and post-processing libraries. Reboot
   and check the board with
   hailortcli fw-control identify.
4. Try vision demos. Install rpicam-apps,
   then run the YOLO object detection, segmentation and pose demos with
   rpicam-hello and the Hailo post-processing files in
   /usr/share/rpi-camera-assets/.
5. Install the Hailo-Ollama server for LLMs. Download
   the Hailo Model Zoo GenAI Debian package from Hailo’s Developer Zone
   (Raspberry Pi’s guide references version 5.1.1) and install it with
   dpkg. Start hailo-ollama, list available
   models at http://localhost:8000/hailo/v1/list, pull one
   (for example qwen2:1.5b), and send prompts to the
   Ollama-style /api/chat endpoint.
6. Optional chat interface. Raspberry Pi’s guide runs
   Open WebUI in Docker, because Open WebUI doesn’t support Python 3.13,
   the version in Raspberry Pi OS Trixie. Point it at the hailo-ollama
   server and open http://127.0.0.1:8080 in a browser.

Because Hailo-Ollama mimics the Ollama REST API, tools like Open
WebUI and LangChain can use it. Hailo’s docs note it handles one
connection at a time, so it isn’t a multi-user server.

Real-world
performance: HAT vs. the Pi’s own CPU

This is where expectations need adjusting. Here’s what independent
testers reported:

• CNX Software (January 20, 2026) ran the same models
  on the HAT through hailo-ollama and on the Pi’s CPU through regular
  Ollama. The CPU was faster every time. For example, it measured
  Qwen2.5-Instruct 1.5B at 6.74 tokens per second on the HAT vs. 11.73 on
  the CPU, DeepSeek R1 1.5B at 6.72 vs. 9.04, and Llama 3.2 3B at 2.60 vs.
  4.78. The closest result was Qwen2.5-Coder 1.5B, at 8.06 vs. 10.26. CNX
  noted the HAT figures may include prompt processing time, so true
  generation speed could be slightly higher.
• Jeff Geerling (January 15, 2026) tested every model
  Hailo had released on an 8GB Pi 5 and wrote that “the Pi’s built-in CPU
  trounces the Hailo 10H,” with Qwen2.5 Coder 1.5B the only near-tie. He
  found the HAT slightly more efficient in most cases, and pointed to
  power as the reason it falls behind: the Pi’s SoC can use up to 10W,
  while the Hailo chip is limited to about 3W.
• A hobbyist benchmark on schwab.sh (January 25,
  2026) reported similar HAT numbers: about 8 tokens per second for Qwen2
  1.5B and 2.65 for Llama 3.2 3B.
• Hailo’s own figures for the current model zoo list
  9.89 tokens per second for Llama3.2-1B, 7.35 for Qwen2.5-1.5B and 4.78
  for Qwen3-1.7B, with time to first token under one second. Hailo doesn’t
  state the host platform for these numbers, so treat them as vendor
  figures.
• Edge-LLM-Bench, an academic paper from an ACM
  MobiSys workshop, found the Hailo-10H didn’t improve token speed over
  the Pi 5 CPU for small models, but reports it cut total system power by
  51% and freed the host CPU and memory.

The upside is real but narrow. CNX Software showed the Pi’s CPU
nearly idle while the HAT generated text, so your Pi can keep running
Home Assistant or your own code at full speed.

The 16GB Pi 5 alternative

If your goal is simply “a capable local LLM on a Raspberry Pi,”
compare the HAT with a Raspberry Pi 5 16GB ($305 list
as of October 23, 2026) running llama.cpp or Ollama on the CPU. You
get:

• Much bigger models. Jeff Geerling ran a compressed
  Qwen3 30B A3B model (squeezed into about 10GB by ByteShape) with
  llama.cpp on a 16GB Pi 5. It built a working to-do list app that the
  tiny models on the HAT couldn’t complete well in his tests. He was clear
  that it’s slow, but it worked.
• Any GGUF model from Hugging Face or the Ollama
  library, with longer context windows, LoRA adapters and the full Ollama
  CLI.
• Faster small models, per the CNX Software and Jeff
  Geerling results above.

Raspberry Pi 5 16GB: Check price on Amazon (paid link)

Jeff Geerling summed up the trade-off: the HAT’s 8GB is “not quite
enough to give this HAT an advantage over just paying for the bigger
16GB Pi with more RAM, which will be more flexible and run models
faster.”

If you already own a 4GB or 8GB Pi 5, the $200 HAT costs less than a
new $305 16GB board, but it only wins if you specifically want to
offload the work or save power.

Caveats and what isn’t
finished

• Running vision and an LLM at the same time is the headline
  feature, but it was unproven at launch. Jeff Geerling got
  segmentation faults and “device not ready” errors when he tried to run
  two models at once, and he found no working examples from Hailo at the
  time. We haven’t found an official Raspberry Pi example of this mixed
  mode in the current docs, so check Hailo’s hailo-apps repository and
  forum before relying on it.
• Early software rough edges. At launch, Jeff
  Geerling found Hailo’s Pi 5 vision examples not yet updated for the
  Hailo-10H, though Raspberry Pi’s rpicam demos worked. Hailo’s model zoo
  now targets HailoRT 5.4.0.
• Small models are small. Raspberry Pi itself says
  edge LLMs sized for the HAT aren’t designed to match the knowledge of
  large cloud models and work best on a constrained task or after
  fine-tuning.
• Version juggling. Raspberry Pi’s guide references
  the 5.1.1 GenAI package, but some newer Hailo models need newer
  runtimes. Match your HailoRT version to the model files.
• One PCIe slot. The HAT takes the Pi 5’s only PCIe
  connector, so plan on booting from microSD or USB storage unless you buy
  a third-party PCIe splitter board.
• Price volatility. Raspberry Pi has said it will
  reverse its memory-driven price increases when memory costs fall, but
  there’s no date for that.

If you’re setting up a Pi as a general Linux box first, our guide to
the best
Linux distros in October 2026 covers the wider landscape, though the
Hailo stack is officially supported only on Raspberry Pi OS.

FAQ

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

No. It connects through the Raspberry Pi 5’s PCIe connector, which
the Pi 4 doesn’t have. Raspberry Pi’s documentation lists a Raspberry Pi
5 with 64-bit Raspberry Pi OS Trixie as the requirement.

What’s the biggest
LLM the AI HAT+ 2 can run?

Hailo’s current model zoo tops out at Qwen3-1.7B for text and 2B for
vision-language models. Raspberry Pi has said models sized for the HAT’s
8GB “typically run at 1–7 billion parameters,” and reviewers ran a Llama
3.2 3B build at launch. Larger models depend on Hailo releasing compiled
versions, because you can’t convert arbitrary models yourself without
Hailo’s compiler.

Can I use regular
Ollama models with it?

No. Hailo-Ollama copies the Ollama REST API, but it only runs models
compiled into Hailo’s HEF format from Hailo’s own list. Standard GGUF
models from the Ollama library run on the Pi’s CPU, not on the HAT.

Is
the AI HAT+ 2 faster than running Ollama on the Pi 5 CPU?

Not for text generation, based on the published reviews. CNX Software
and Jeff Geerling both measured the Pi 5 CPU generating tokens faster
than the HAT for nearly every model. The HAT’s advantages are freeing
the CPU and RAM, and lower power use.

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

$200 on Raspberry Pi’s official product page as of October 23, 2026.
It launched at $130 on January 15, 2026, and rose by $50 on April 1,
2026 due to memory costs. Reseller prices and local taxes vary.

Should I buy the
original AI HAT+ instead?

If you only need camera-based vision such as object detection or pose
estimation, yes. The original AI HAT+ starts at $70 and Raspberry Pi
says the HAT+ 2’s vision performance is broadly similar to the 26 TOPS
version. Only the HAT+ 2 can run LLMs, VLMs and Whisper.

Sources

• Raspberry Pi AI HAT+ 2 product page (specs, $200 price, production
  lifetime): https://www.raspberrypi.com/products/ai-hat-plus-2/
• Raspberry Pi news, “Introducing the Raspberry Pi AI HAT+ 2” (Jan 15,
  2026):
  https://www.raspberrypi.com/news/introducing-the-raspberry-pi-ai-hat-plus-2-generative-ai-on-raspberry-pi-5/
• Raspberry Pi news, “A new 3GB Raspberry Pi 4 for $83.75, and more
  memory-driven price increases” (Apr 1, 2026):
  https://www.raspberrypi.com/news/a-new-3gb-raspberry-pi-4-for-83-75-and-more-memory-driven-price-increases/
• Raspberry Pi documentation, AI software (setup, hailo-h10-all,
  hailo-ollama, Open WebUI):
  https://www.raspberrypi.com/documentation/computers/ai.html
• Raspberry Pi documentation, AI HATs:
  https://www.raspberrypi.com/documentation/accessories/ai-hat-plus.html
• Raspberry Pi AI HAT+ product page:
  https://www.raspberrypi.com/products/ai-hat/
• Raspberry Pi 5 product page (16GB price):
  https://www.raspberrypi.com/products/raspberry-pi-5/
• Hailo Model Zoo GenAI, supported models:
  https://github.com/hailo-ai/hailo_model_zoo_genai/blob/main/docs/MODELS.rst
• Hailo Model Zoo GenAI, usage and Hailo-Ollama limits:
  https://github.com/hailo-ai/hailo_model_zoo_genai/blob/main/docs/USAGE.rst
• Hailo apps repository (VLM and GenAI examples):
  https://github.com/hailo-ai/hailo-apps
• Hailo Community forum, Hailo-Ollama model support (May 2026):
  https://community.hailo.ai/t/hailo-ollama-support-for-model-explorer-genai-models/19275
• Jeff Geerling, “Raspberry Pi’s new AI HAT adds 8GB of RAM for local
  LLMs” (Jan 15, 2026):
  https://www.jeffgeerling.com/blog/2026/raspberry-pi-ai-hat-2/
• CNX Software, Raspberry Pi AI HAT+ 2 review (Jan 20, 2026):
  https://www.cnx-software.com/2026/01/20/raspberry-pi-ai-hat-2-review-a-40-tops-ai-accelerator-tested-with-computer-vision-llm-and-vlm-workloads/
• CNX Software, April 2026 Raspberry Pi price increases with resulting
  prices:
  https://www.cnx-software.com/2026/04/01/raspberry-pi-4-3gb-launched-for-83-75-further-price-increases-announced-across-the-board-for-4gb-ram-hardware/
• schwab.sh, “Benchmarking the Raspberry Pi AI HAT+ 2 for Local LLMs”:
  https://www.schwab.sh/blog/hailo-ai-hat-benchmarks/
• Edge-LLM-Bench (ACM MobiSys 2026 workshops):
  https://dl.acm.org/doi/10.1145/3812836.3814998
• The Verge, launch coverage (Jan 15, 2026):
  https://www.theverge.com/news/862748/raspberry-pi-ai-hat-2-gen-ai-ram

---

Last updated: October 23, 2026