It is built on the Simple Attention Network, compressed to CQ2-bit with Cactus Quants, and packaged in Cactus's own engine. The browser demo runs the model in WebAssembly and shows tool routing, parallel calls, multi-step actions, extraction, classification, and refusal behavior.
Needle 2 is useful for offline assistants, smart-home control, phones, wearables, robots, and embedded interfaces where memory, latency, and connectivity are constrained. Cactus reports up to 500 tokens per second on a Raspberry Pi 5 and high throughput on supported XR devices and phones.
