Cactus Compute

Needle 2 is a 45M parameter open model that ships as one 14 MB binary, aimed at the cheap end of the edge: sub $200 phones, wearables, smart home gear, small robots, even microcontrollers with no GPU and no NPU. After CQ2 bit compression a whole session fits in roughly 28 MB of RAM, by the makers' own measurement. It is trained for what those devices actually do rather than for general chat, which means calling tools, driving device actions and pulling structured fields out of messy text. A playground on the site loads the model in WebAssembly inside the browser, so the smart home, robot and form filling examples run without installing anything. Cactus reports decode speeds in the hundreds of tokens per second on hardware like a Raspberry Pi 5, and the weights sit on Hugging Face under Apache 2.0.