Raspberry Pi 5 gets LLM smarts with AI HAT+ 2
- Reference: 1768477771
- News link: https://www.theregister.co.uk/2026/01/15/pi_5_ai_hat_2/
- Source link:
On paper, the specifications look great. The [1]AI HAT+ 2 delivers 40 TOPS (INT4) of inference performance. The Hailo-10H silicon is designed to accelerate large language models (LLMs), vision language models (VLMs), and "other generative AI applications."
Computer vision performance is roughly on a par with the 26 TOPS (INT4) of the previous AI HAT+ model.
[2]
These components and 8 GB of onboard RAM should take a load off the hosting Pi, so if you need an AI coprocessor, you don't need to blow through the Pi's memory (although more on that later).
[3]
[4]
The hardware plugs into the Pi's GPIO connector (we used an 8 GB Pi 5 to try it out) and communicates via the computer's PCIe interface, just like its predecessor. It comes with an "optional" passive heatsink – you'll certainly need some cooling solution since the chips run hot. There are also spacers and screws to fit the board to a Raspberry Pi 5 with the company's active cooler installed.
[5]
AI HAT+ 2 on Raspberry Pi 5
Running it is a simple case of grabbing a fresh copy of the Raspberry Pi OS and installing the necessary software components. The AI hardware is natively supported by rpicam-apps applications.
In use, it worked well. We used a combination of Docker and the hailo-ollama server, running the Qwen2 model, and encountered no issues running locally on the Pi.
However, while 8 GB of onboard RAM makes for a nice headline feature, it seems a little weedy considering the [6]voracious appetite AI applications have for memory. In addition, it is possible to specify a [7]Pi 5 with 16 GB RAM for a price.
[8]New York's incoming mayor bans Raspberry Pi at his inauguration party
[9]Cheaper 1 GB Raspberry Pi 5 lands as memory costs go through the roof
[10]Struggling to heat your home? How about 500 Raspberry Pi units?
[11]Raspberry Pi prices hiked as AI gobbles all the memory
And then there's the computer vision, which is broadly the same 26 TOPS (INT4) as the earlier AI HAT+. For users with vision processing use cases, it's hard to recommend the $130 AI HAT+ 2 over the existing AI HAT+ or even the [12]$70 AI camera .
Where LLM workloads are needed, the RAM on the AI HAT+ 2 board will ease the load (although simply buying a Pi with more memory is an option worth exploring). According to Raspberry Pi, DeepSeek-R10-Distill, Llama3.2, Qwen2.5-Coder, Qwen2.5-Instruct, and Qwen2 will be available at launch. All (except Llama3.2) are 1.5-billion-parameter models, and the company said there will be larger models in future updates.
[13]
The size compares poorly with what the cloud giants are running (Raspberry Pi admits "cloud-based LLMs from OpenAI, Meta, and Anthropic range from 500 billion to 2 trillion parameters"). Still, given the device's edge-based ambitions, the models work well within the hardware constraints.
This brings us to the question of who this hardware is for. Industry use cases that require only computer vision can get by with the previous 26 TOPS AI HAT+. However, for tasks that require an LLM or other generative AI functionality but need to keep processing local, the AI HAT+ 2 may be worth considering. ®
Get our [14]Tech Resources
[1] https://www.raspberrypi.com/news/introducing-the-raspberry-pi-ai-hat-plus-2-generative-ai-on-raspberry-pi-5/
[2] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=2&c=2aWlxmf2A38S0UGJNH_lPxAAAA1g&t=ct%3Dns%26unitnum%3D2%26raptor%3Dcondor%26pos%3Dtop%26test%3D0
[3] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44aWlxmf2A38S0UGJNH_lPxAAAA1g&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[4] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33aWlxmf2A38S0UGJNH_lPxAAAA1g&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[5] https://regmedia.co.uk/2026/01/15/ai_hat_on_pi.jpg
[6] https://www.theregister.com/2024/08/25/ai_pc_buying_guide/
[7] https://www.theregister.com/2025/01/09/the_ultimate_pi_5_arrives/
[8] https://www.theregister.com/2025/12/31/zohran_mamdani_raspberry_pi_ban/
[9] https://www.theregister.com/2025/12/01/raspberry_pi_5_1gb/
[10] https://www.theregister.com/2025/10/03/thermify_heathub_raspberry_pi/
[11] https://www.theregister.com/2025/10/01/raspberry_pi_price_hikes/
[12] https://www.raspberrypi.com/news/raspberry-pi-ai-camera-on-sale-now/
[13] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44aWlxmf2A38S0UGJNH_lPxAAAA1g&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[14] https://whitepapers.theregister.com/
Optionally on top, rather than in.
Robots, obviously
For as long as RPi has existed people have tried building robots with it, trying various ways to add vision, autonomy, and natural language interaction. This is the obvious target audience for this. Not people who want to play with chatbots, "agentic AI" (whatever that means this week) and so on. That is to say, it's for people who actually want to mostly do machine learning, the thing that's now been rebranded, thrown together with tangentially related stuff, diluted with IP theft, and vilified as "AI".
Re: Robots, obviously
And lets be honest, people care now about having things hosted at home away from the internet/cloud/someone-elses-computer. If they want an LLM for simple shit, this is ideal. And maybe possibly more cost effective than finding a PC with a good enough GPU/Graphics card included with it.
Re: Robots, obviously
This, definitely! It's slowly becoming a trend, and for good reason. It is quite annoying (and scary) to have a smart home solution that is hosted somewhere online - in the future we will have more and more solutions that rely on local information processing and storage. This kind of AI HAT enables the creation of a fantastic smart home solution that could easily be integrated with some of the existing offline packages.
Smart home management kit that is COMPLETELY separated from the Internet - that's is the future I would like to live in (without big brothers in every room - they should be limited to phones, and that's it)
Re: Robots, obviously
It is quite annoying (and scary) to have a smart home solution that is hosted somewhere online
I am all for off-line processing, escaping the tentacles of the cloud. Though I am not so enthusiastic about throwing off-line AI into the mix.
Who it's for? What it has always been for. It's a learning device for people who want to mess about and try stuff. The Hailo 8 couldn't do LLMs very well. The 10h can and has 5 models ready-cooked, though relatively speaking they are SLMs.
For all the crying, wailing and toys out of pram tantrum comments on Register forums, AI isn't going anywhere, it will only increase and people want to learn about it. Raspberry Pi can't afford to be left behind.
Better be quick and pick up a bargain, before prices _really_ start hitting the roof: https://shop.pimoroni.com/products/raspberry-pi-500-plus?variant=55744218202491
Maybe they will just cannibalize these machines for memory to make more a1 "hats" . . .
Just because people disagree with you on the future of AI, that doesn't make their comments "toys out of pram tantrums", when they clearly weren't.
https://www.wheresyoured.at/wheres-the-money/
I for one would prefer if some effort went into being able to use the hardware without the full desktop environment dependencies, "AI" or no "AI". Everything and their cat seems to require Wayland (or X) and SystemD nowadays, and not many fucks appear to be given by Debian or the Pi foundation. To me the Pi was always meant to shine as a minimalistic low power platform for things like automation, robotics, IoT, signage, media players, etc, etc - none of which benefit in any way from the added bloat of a desktop. I found it tremendously useful as a cross between a microcontroller and a full blown Linux system. If you really need the kind of muscle provided by the Pi 5 (gaming? web browsing? transcoding?), why not opt for one of the many mini PCs made by the likes of Lenovo, Dell and HP instead? I recently picked up a HP EliteDesk with a Ryzen 5 CPU, Vega iGPU, 16 GB RAM, NVMe and SATA - an order of magnitude better performance for about the same $$$, in a sturdy & practical case with standard size connectors to boot. I personally do not see any application for all those MIPS which simultaneously requires the I/O capabilities of the Pi platform. If I dared to dream it would be great if the Pi 3 / Zero platform could be kept active and gain a wide voltage range (say 5-28V) power input jack that didn't require USB. But maybe that's just me.
I run NextCloud over Devuan on a Pi. Headless - no Wayland, X11 or desktop. No systemd. OS on NVME, data on HDD mirrored with LVM2. Perfectly doable.
Try setting one up as a webcam.
No interest in that. Somewhere at the back of the queue is the idea of using one to replace the existing optics/processing in a slide convertor.
Try setting one up as a webcam.
Are you saying there are problems doing that?
FWIW I've recently finished a project where two Pi5s were used headless with two High Quality Picams to take timelapse over a period of about six months. Images saved both to a network share and locally. All controlled in Python with the standard libraries, power on boots straight into the script and maybe three quarters of the script deals with naming the images, sorting them into appropriate folders per day, changing the cadence according to the time of day and calendar date. The hardest part was working out why it would freeze after (IIRC) 1023 images, not getting to the root of the problem but realising it was some kind of resource leak and re-writing the code to close and re-open the camera before that limit was reached.
The "lite" version of the OS is your friend and other Pi-compatible OSes have similar options I believe.
M.
The attraction of the Rasp Pi (and Linux in general) is its flexibility.
You CAN run the unit headless, or you can run a desktop for your child to learn various computing. It does both.
Accept that what YOU do is not the gold standard for everyone and that me doing my own thing does not take away from what you do in the slightest.
What's with the aggro - we seem to agree 100%?
Judging from the downvotes my post seems to have been misunderstood. When I talk about "the hardware" I'm specifically referring to things unique to the Pi; the camera and display interfaces, the GPIO, quirks of the SoC. Of course you can run whatever Linux distro you like, and customise it to your liking, but when you need to support the Pi specific hardware you often have to install packages which depend on Wayland and/or SystemD. It may be wrong to blame the Pi Foundation for this - but I cannot think of anyone who would be better positioned to do something about it. I also reject the notion that the best answer to bloat is to increase system resources.
"you often have to install packages which depend on Wayland and/or SystemD"
This is obviously the choice of whoever wrote the packages and having the Pi Foundation mandate otherwise is very much contrary to the idea of free software.
Having said that Devuan puts a directory in /var that a number of applications expect and which seems to keep them happy. If that's insufficient you could check to see if there are Devuan variants of the libraries.
Likewise, it's worth checking to see if there are X11 variants of packages that expect Wayland. For instance on Trixie/Excalibur Pipewire kept logging error messages on an X11 session until I discovered that there was a library for Pipewire under X11 that wasn't installed by default.
If a third party package requires the desktop to run (Official Pi camera applications etc do not ), why would the Pi people be able to do anything about it?
1. It's NOT the Pi Foundation. They are completely independant charity who do NOT do the HW/SW that people think of as Pi,. they just do the educational charity stuff. Raspberry Pi Ltd do the HW/SW that people use.
2. PiOS can be installed without any desktop - PiOS Lite. No dependencies on X/Wayland. So plenty of fucks given. The same applies to many other Linux OS that run on the Pi.
3. Putting the circuitry to support 5-28V would make all the boards much bigger and hotter. Not going to happen in the current form factors. Fitting everything on is hard enough as it is.
4. It IS a full-blown Linux system.
I for one welcome our raspberry PIE overlords!!
FFS, they will try and shoehorn AI into bloody anything these days.