tl;dr: Cactus has released Needle 2, an open-source (Apache 2.0) 45M-parameter model for tool calling and structured extraction, shipped as a 14MB binary with a 28MB RAM ceiling that runs on microcontrollers, cheap phones, and Raspberry Pis at 300–1,500 tokens/sec. It uses a custom Simple Attention Network architecture with 2-bit quantization baked into training (not post-hoc) and a Hadamard MLP plus hashed n-gram engram tables to cut FLOPs. On function-calling benchmarks like Mobile Actions and Seal-Tools, it trades wins with FunctionGemma 270M, LFM2.5 230M, and Apple FM despite being 5–70× smaller.