A free AI model just landed that is smaller than a decent phone photo, runs on a $35 Raspberry Pi, and never sends a byte to the cloud. If you have been waiting for AI to get cheap and simple enough for a counter device or a back-office tablet, this is worth five minutes of your time.
What's happening
Cactus Compute, a small edge-AI startup, released Needle 2 this week. It is a 45-million-parameter model that ships as a single 14MB binary and runs a full session using only 28MB of RAM. That is a remarkable number: most AI models you interact with today pull gigabytes of compute from cloud servers. Needle 2 runs the whole thing on your own hardware, start to finish.
The open-source release is under Apache 2.0, so it is free for commercial use with no licensing fees and no usage caps. The GitHub repository has a quick-start guide you can follow in under ten minutes.
What it actually does is the important part. Needle 2 is built for structured tasks: extracting fields from receipts, routing a customer request into the right category, filling in a form from a photo or spoken input, pulling answers from a knowledge base. On a Raspberry Pi 5, it handles those tasks at 500 tokens per second. On a cheap Android phone, similar speed. Cactus specifically names retail kiosks, point-of-sale devices, and invoice extraction as target use cases because those are exactly the repetitive data jobs that run on modest hardware at a counter or front desk.
Try this this week
- Read the Applications section at cactuscompute.com/needle. The examples for receipt extraction and form processing are concrete and close to what a small shop or clinic actually deals with.
- Write down one manual data task at your place of business: receipt amounts you type into a spreadsheet, intake form answers you re-enter, FAQ questions you type the same reply to every day. That is your starting test case.
- Try it on your laptop first. Any modern laptop hosts Needle 2 without a GPU. The download takes seconds. Pull the repo, run the quick-start, and paste in a sample receipt or form to see what it extracts.
- If you have a Raspberry Pi collecting dust, it is a perfect permanent home for this. Pi 5 at full speed, Pi 4 works fine too.
- Think about the kiosk angle. A cheap tablet or Pi at your counter, running Needle 2 locally, could handle intake questions, log receipt data, or answer stock queries. No subscription, no internet dependency.
The bigger picture
Cloud AI is convenient, but the bill grows with every use and your data travels somewhere you did not choose. What is changing right now is that AI is moving off the cloud and into the device itself. Needle 2 is one of the first models small enough to embed directly in a counter kiosk or a cheap tablet, doing one specific job quietly and offline. For small shops and nonprofits, that is a meaningful shift: the tool stays yours, the data stays local, and the cost stays zero. The smallest working thing sometimes fits in 14 megabytes.