On 2026-09-22, Raspberry Pi News covered Cactus Compute’s Needle 2: a ~14MB function-calling model that turns plain English into local actions on a Raspberry Pi 5 using the CPU alone (no dedicated AI HAT) (RPi News).
Needle is an action-only LLM, not a chatbot. You declare Python functions; the model selects a tool and fills arguments. Off-topic prompts (for example “What is the capital of France?”) return an empty function_calls: [].
Size, license, install
Vendor + RPi + HF card agree on an open ~45M-parameter model shipped as a single ~14MB binary; native session ~28MB RAM (RPi Python demo process peak 43–46.4MB including interpreter). Weights and code: Hugging Face, GitHub—both Apache-2.0. Install path in the post: pip install cactus-needle.
Latency (RPi demo, not desk-measured)
The latency figures in the RPi post are an author demo on Raspberry Pi 5, 8GB, Raspberry Pi OS, CPU only, cactus-needle 2.0.7—not measured by this desk (RPi News). Wall-clock complete() before Python tool runs:
| Prompt (examples) | Wall-clock |
|---|---|
| “Turn the LED on” | 78 ms |
| “How hot is this Raspberry Pi?” | 149 ms |
| “Blink the LED 2 times” | 83 ms |
| “Take a photo” | 76 ms |
| “Save a note…” | 107 ms |
| Off-topic capital | 92 ms |
In that same table, prefill ~461–488 tok/s and decode ~248–314 tok/s. Separately, Cactus materials cite decode up to ~500 tok/s on Pi 5—vendor-reported. That figure is not the same as the RPi table’s ~248–314 decode range.
Who should care
Edge / GPIO teams who want a tiny Apache-2.0 action router on Pi-class hardware should read the RPi post and the HF card. Treat it as structured tool selection, not general chat.
The Campfire
CommentsThanks. Your comment is waiting for the desk. Once it is approved it appears here within about a minute.
Add a comment