The Raspberry Pi AI Kit Is Discontinued. Here Are Your Real Options.

Raspberry Pi AI Kit migration guide

The Raspberry Pi AI Kit Is Discontinued. Here Are Your Real Options.

Last updated: 21 October 2026
Short answer: Yes, the Raspberry Pi AI Kit is discontinued. Raspberry Pi says it is “no longer in production” and points buyers to the Raspberry Pi AI HAT+. The Kit’s Hailo-8L chip is not discontinued: it continues in the 13 TOPS Raspberry Pi AI HAT+.

If you own an AI Kit, you have three realistic options:

  1. Move to the Raspberry Pi AI HAT+ and keep your Hailo models.
  2. Move to the Sixfab AI HAT+ (DEEPX) and recompile from ONNX.
  3. Buy nothing, and keep your AI Kit running by pinning its software versions.

The Raspberry Pi AI Kit was the $70 Hailo-8L bundle that introduced a lot of people to on-device AI in 2024. If you own one, or you were about to build on one, this page lays out your options honestly, including the one where you buy nothing.

Straight answers

Three paths. None of them is right for everyone.

Shortest migration
Stay with Hailo

Use the Raspberry Pi AI HAT+ if your HEF models and Hailo pipeline already work.

New build / second source
Move to DEEPX

Use the Sixfab AI HAT+ for a new build or a second NPU vendor. YOLO models export in one command.

No new requirements
Buy nothing

Keep the discontinued AI Kit running by freezing the software stack that already works.

What happened

What actually happened

The AI Kit launched in June 2024. It was a Hailo-8L accelerator (13 TOPS) on an M.2 2242 card, pre-installed in the M.2 HAT+, and it worked with the Raspberry Pi 5 only.

In late 2024 Raspberry Pi put the same silicon family onto a single board, the Raspberry Pi AI HAT+, and the Kit was retired. Raspberry Pi’s own product page now says the Kit is no longer in production. Retailers such as The Pi Hut mark it discontinued and redirect buyers to the Raspberry Pi AI HAT+.

Two other things happened around it, and they matter more than the discontinuation itself.

The software moved

Raspberry Pi OS moved to Trixie, HailoRT moved through its 4.x chain (4.23 was the latest release as of this writing), and the kernel moved to 6.12.

Popular AI Kit tutorials, including the most-viewed ones, pin versions from 2024 that no longer install cleanly. If your setup broke, it probably wasn’t your fault.

Hailo is now owned by Microchip

Microchip Technology signed a definitive agreement to acquire Hailo in July 2026 and completed the deal on 21 September 2026.

Microchip says it plans to keep supporting Hailo’s existing products, software and customers. That’s a reassuring statement, but it isn’t a product roadmap, and none has been published yet.

Factor it into any multi-year hardware decision, and don’t read more into it than that.

Software reality

Why your old tutorial broke anyway

Even before the discontinuation, AI Kit owners were fighting version drift, and the public forums show it:

Trixie upgradesTrixie upgrade threads where the Hailo driver stops loading.
Kernel 6.12DKMS builds failing against kernel 6.12.
Python 3.13“HailoRT not detected” errors on Python 3.13, even when the device itself works.
The practical lessonPick your OS image, runtime version and model format deliberately. Pin them and upgrade on your own schedule, not by accident.

The lesson applies to every option below. On-device AI software stacks are young.

Your three options

Choose the path that matches the system you actually have

1
Official continuation

Raspberry Pi AI HAT+

Keep Hailo, keep the same model toolchain, and carry existing HEF models forward.

See Option 1 →
2
Alternative NPU path

Sixfab AI HAT+ with DEEPX

Recompile from ONNX, use the DEEPX runtime, and add a second NPU vendor to your hardware strategy.

See Option 2 →
3
No purchase required

Keep your AI Kit running

Freeze the working software stack and keep using the hardware you already own.

See Option 3 →
1
Stay on Hailo

Option 1: Stay on Hailo with the Raspberry Pi AI HAT+

This is the official continuation path. The Raspberry Pi AI HAT+ comes in two versions: 13 TOPS (Hailo-8L, from $70) and 26 TOPS (Hailo-8). It integrates with the Raspberry Pi camera software, and Raspberry Pi has committed to keeping it in production until at least January 2030.

Choose this if: your existing Hailo models (HEF files) and pipelines work today and you want the shortest possible migration. Your compiled models carry over and your tools stay the same.
Go in with open eyes:

You keep the same Hailo Dataflow Compiler toolchain, the same need to manage versions carefully that broke the tutorials, and a vendor that has just changed owners.

Microchip has said it will keep supporting the existing product line. The longer-term roadmap under the new owner hasn’t been published yet.

Need generative AI rather than vision?

Raspberry Pi also sells the Raspberry Pi AI HAT+ 2: Hailo-10H, 40 TOPS (INT4), 8 GB of on-board RAM, $200, and in production until at least January 2036.

It’s built for running LLMs and vision-language models locally. It uses a different chip from the AI Kit’s Hailo-8L, so confirm with Hailo that your compiled models run on it before you assume they carry over.

2
Change NPU vendor

Option 2: Move to DEEPX with the Sixfab AI HAT+

This is the path Sixfab builds for. Here’s what it involves.

The Sixfab AI HAT+ puts a DEEPX NPU on the Raspberry Pi 5 in two versions: 13 TOPS with 1 GB of dedicated NPU memory ($63), or 25 TOPS with 2 GB ($95). The NPU itself draws roughly 3 W typically. You install it through APT from Sixfab’s signed repository, and it has Python and C++ APIs.

The 25 TOPS version is available now. The 13 TOPS version ships in early 2027.

Here’s what migration really means

Your Hailo HEF files do not transfer. You go back to your original model (ONNX or a YOLO checkpoint) and compile it for DEEPX.

1. Native Ultralytics exportUltralytics ships a native DEEPX export (format="deepx"), so a YOLO model is one command away. The export runs on an x86-64 Linux machine, not on the Pi itself. The Pi then runs the compiled model.

Read the Ultralytics DEEPX integration →

2. Published performanceOn the 25 TOPS version, our benchmark measured 50.1 FPS on YOLOv8l and 178.7 FPS on YOLO26n in async NPU mode. The DEEPX vs Hailo benchmark puts them next to Hailo’s, including what PCIe Gen3 ×1 does and doesn’t cost you.

See the DEEPX vs Hailo benchmark →

Frigate NVR: current status

We wrote a DEEPX detector integration for Frigate, and the pull request is open upstream. You can run it today by installing from our repository.

It becomes official when the Frigate maintainers merge it. We will update this page the day that changes.

Choose this if: you’re starting a new build, you want a competitive price per TOPS and a direct Ultralytics-to-NPU workflow, or you want a second source so your product isn’t tied to one NPU vendor.

Not sure your model will port? Send it to us and we’ll tell you whether it ports cleanly before you spend anything.

3
The zero-cost path

Option 3: Buy nothing and keep your AI Kit running

It’s unfashionable, but it’s a valid choice. Discontinued doesn’t mean disabled. Your AI Kit keeps working for as long as you control its software.

How to do it

Stay on your current known-good OS image.
Pin the HailoRT version that matches your compiled models.
Turn off unattended upgrades for the kernel and the Hailo packages.
Back up the whole SD card as an image.
A frozen, working 13 TOPS setup beats a broken upgraded one.

Choose this if: your project is finished and running, and you don’t need more performance or new models. Come back to this page when you actually need to retrain or rebuild.

Side-by-side comparison

Raspberry Pi AI Kit vs Raspberry Pi AI HAT+ vs Sixfab AI HAT+

SIXFAB · Raspberry Pi AI Kit Migration GuideComparison · Updated 21 Oct 2026
ComparisonRaspberry Pi AI KitRaspberry Pi AI HAT+Sixfab AI HAT+
StatusDiscontinuedIn production, committed to at least Jan 203025 TOPS in stock; 13 TOPS ships early 2027
NPUHailo-8LHailo-8L / Hailo-8DEEPX DX-M1ML / DX-M1M
Performance13 TOPS13 / 26 TOPS13 / 25 TOPS
Dedicated NPU memoryNoneNone1 GB / 2 GB
PriceWas ~$70From $70 (13 TOPS)$63 / $95
Model workflowONNX or TFLite → HEF (Hailo Dataflow Compiler)Same as AI Kit; existing HEF models carry overONNX → DXNN (DX-COM); native Ultralytics export
Camera stackrpicam integrationrpicam integrationPython / C++ SDK, GStreamer pipeline
Frigate NVRCommunity-supported (Hailo detector)Community-supported (Hailo detector)Integration PR open; installable from Sixfab repo today
Vendor situationHailo, acquired by Microchip (closed Sep 2026)SameDEEPX, independent; sold by Sixfab, an official Raspberry Pi Design Partner
© Sixfab · Migration comparisonsixfab.com/raspberry-pi-ai-kit-discontinued

Not sure which path fits?

Tell us what you run today: board, model and pipeline. We’ll give you a straight answer, even if the answer is “stay where you are.”

Questions

FAQ

Is the Raspberry Pi AI Kit discontinued?

Yes. Raspberry Pi says the AI Kit is no longer in production, and major resellers list it as discontinued. Raspberry Pi’s successor product is the Raspberry Pi AI HAT+.

Is the Hailo-8L discontinued?

No. The AI Kit product is discontinued, but the Hailo-8L chip it used continues in the 13 TOPS Raspberry Pi AI HAT+, which Raspberry Pi has committed to producing until at least January 2030.

What replaced the Raspberry Pi AI Kit?

The Raspberry Pi AI HAT+, a single-board design with the same Hailo silicon family in 13 TOPS and 26 TOPS versions. For generative AI workloads, Raspberry Pi also sells the Raspberry Pi AI HAT+ 2 (Hailo-10H, 40 TOPS INT4). If you’re open to changing NPU vendor, the Sixfab AI HAT+ is a DEEPX-based alternative with native Ultralytics export.

What is the difference between the Raspberry Pi AI Kit and the Raspberry Pi AI HAT+?

The AI Kit was a Hailo-8L M.2 module mounted on a separate M.2 HAT+. The Raspberry Pi AI HAT+ puts the Hailo chip directly on the HAT board and adds a 26 TOPS Hailo-8 version alongside the 13 TOPS Hailo-8L one. Both work with the Raspberry Pi 5 only.

Will my existing AI Kit keep working?

Yes, as long as you control its software. Pin your OS, kernel and HailoRT versions to a known-good combination and turn off automatic upgrades for those packages. Most breakage reported on the forums comes from version drift, not hardware failure.

Is Hailo still supported after the Microchip acquisition?

Microchip completed its acquisition of Hailo in September 2026 and says it plans to continue supporting Hailo’s existing product portfolio, software environment and customers. A long-term product roadmap under Microchip hasn’t been published yet.

Can I reuse my Hailo models on a DEEPX accelerator?

Not directly. Compiled HEF files are Hailo-specific. You recompile from your original ONNX or YOLO model. For YOLO, Ultralytics has a native DEEPX export that makes this a one-command step on an x86-64 Linux machine. Sixfab will check your model’s portability for free before you buy.

Which one should I buy in 2026?

For the shortest migration with existing Hailo models, buy the Raspberry Pi AI HAT+. For a new build, a competitive price per TOPS, a direct Ultralytics workflow, or a second NPU vendor, buy the Sixfab AI HAT+. If your system is running and has no new requirements, keep your AI Kit and freeze its versions.

Migration review

Bring us the model and the pipeline you have today.

We’ll tell you whether the shortest path is Raspberry Pi’s Hailo continuation, a DEEPX migration, or simply leaving a working AI Kit alone.