Cactus launches a 14MB tool-calling model for tiny devices

Cactus released Needle 2, an open 45M-parameter model for tool calling, device use and structured extraction. The model ships as a single 14MB binary, runs in 28MB of RAM, and is built on the company’s Simple Attention Network with CQ2-bit compression and a dedicated engine.

Needle 2 targets devices without a GPU or NPU, including budget phones, Raspberry Pis, microcontrollers, wearables, small robots and connected home hardware. Cactus says it reaches 500 tokens/sec decode speed on a Raspberry Pi 5, 400-1,500 tokens/sec on VR devices like Meta Quest 3S and Apple Vision Pro, and 300-700 on sub-$200 phones.

The model focuses on mapping user requests to typed functions and schemas rather than general chat. It treats extraction as tool calling, compiles grammar from schemas to prevent malformed JSON, returns confidence scores, and can refuse off-topic requests with an empty call so products can retry or escalate to the cloud.

Benchmark results show Needle 2 trading wins with small on-device models such as FunctionGemma 270M, LFM2.5 230M and Apple FM, while being 5× to 70× smaller and running at 2 bits against their f16. Pebble is using Cactus Needle locally in the Index 01 app for spoken actions without depending on a network connection.