Sam Witteveen · 2026-07-12 · notable
Sam Witteveen: 'Cactus Needle — The 26M Function Calling Model'
Sam Witteveen walks through Cactus Compute's Needle, a 26M-parameter tool-calling model distilled from Gemini 3.1 Flash Lite. The video explains the Simple Attention Network architecture and how the model runs on phones and laptops.
A hands-on video breakdown of Cactus Needle, a 26M-parameter tool-calling model that fits in 14 MB.
What is it?
Sam Witteveen's July 12 video 'Cactus Needle — The 26M Function Calling Model' walks through Cactus Compute's Needle model. Needle is a 26M-parameter open-weight model distilled from Gemini 3.1 Flash Lite for single-shot function calling, quantized to a 14 MB INT4 footprint.
How does it work?
The video explains Cactus's Simple Attention Network (SAN) design — a transformer-style stack with the feed-forward MLP layers removed — and shows the model producing JSON tool calls from natural-language queries. Sam runs it against the sample tasks Cactus published to compare speed and accuracy against FunctionGemma-270M and Qwen-0.6B.
Why does it matter?
Sam Witteveen's channel is a common first-look source for AI engineers evaluating new open-weight models. His breakdown lands as more agent builders look at moving tool-calling loops onto phones and edge devices to avoid API latency and cost.
Who is it for?
Engineers building on-device agent loops and anyone tracking small, specialized function-calling models.
Try it
Watch on YouTube: https://www.youtube.com/watch?v=tt9UJ0NiOzU