Could a palm-sized add-on board really run a large language model on your Raspberry Pi 5 — no cloud, no $3,500 AI box required?
That question anchors this week’s Python on Microcontrollers newsletter from Adafruit, and the answer is a surprisingly confident yes. Two new releases are pushing serious generative AI to the edge. The official Raspberry Pi AI HAT+ 2 turns a Pi 5 into a local inference machine, while NVIDIA’s Jetson Orin Nano Super Developer Kit goes a step further for makers who want maximum horsepower in a compact footprint. And if that weren’t enough, there’s an open-source CubeSat kit that runs MicroPython on the RP2350.
So how much AI can these boards actually handle?
The AI HAT+ 2 pairs a Hailo-10H accelerator with 8GB of dedicated LPDDR4X memory over PCIe, delivering up to 40 TOPS of INT4 inference. That’s enough for local vision models and small generative workloads — think on-device assistants, image captioning, or smart camera projects that never send a frame to the cloud. The Jetson Orin Nano Super raises the ceiling to 67 TOPS, comfortably running vision transformers, vision-language models, and compact LLMs.
The other highlights lean hardware-hacker:
- VyomSat — a high-fidelity, MIT-licensed PocketQube/CubeSat learning kit built on the RP2350 and programmed in MicroPython. Power systems, attitude control, real satellite engineering — for under $200.
- RV Circuit Studio — a new Apache-licensed CircuitPython editor gunning to replace Mu, with board auto-detection, a source-level debugger, serial plotter, and library manager.
- The Replay badge — conference swag that turned out to be a hackable ESP32-S3 computer with an OLED, LED matrix, and MicroPython (yes, it runs DOOM).
Try it yourself
If edge AI has been on your project list, this is the cheapest entry point yet: a Raspberry Pi 5, the AI HAT+ 2, and a weekend. Start with a vision model before jumping to LLMs, and keep a decent power supply handy — accelerators are hungry. You’ll find Raspberry Pi boards, power supplies, and sensors to round out the build at Circuit.Rocks. And if you’d rather aim higher — literally — that $200 MicroPython CubeSat is one of the most ambitious STEM kits we’ve seen all year.
Frequently Asked Questions
What makes the Raspberry Pi AI HAT+ 2 different from earlier AI HATs?
It’s the first official generative-AI add-on for the Pi 5: a Hailo-10H accelerator with 8GB of dedicated LPDDR4X memory delivering up to 40 TOPS of INT4 inference, enough for local vision and small generative AI models rather than just object detection.
How much does the VyomSat CubeSat kit cost and what runs it?
The open-source PocketQube/CubeSat learning kit comes in under $200. It’s built on the Raspberry Pi RP2350 microcontroller, programmed entirely in MicroPython, and MIT licensed, covering power systems through attitude control.
What will I learn if I build an edge AI project like this?
You’ll learn how to deploy and quantize AI models on constrained hardware, work with PCIe accelerators on the Pi 5, manage power budgets, and connect Python code to real sensors and cameras — skills that transfer directly to robotics, IoT, and embedded engineering.
