Needle

Needle 2 is an open 45M-parameter model for tool calling, device use and structured extraction, and the whole model is a single 14MB binary that runs a session in roughly 28MB of RAM. Weights are baked into the engine, so there are no separate model files to manage and inference never touches the network.

Text goes in, JSON comes out. A byte-level grammar compiled from your tool schemas constrains every token, and each response carries a calibrated confidence score. The makers say they built it on their Simple Attention Network work and compressed it to CQ2-bit with Cactus Quants; on their own benchmarks it trades wins with FunctionGemma 270M, LFM2.5 230M and Apple FM at 5x to 70x smaller. This repository is the Python package covering inference, LoRA fine-tuning and export, installed with pip install cactus-needle.