Cactus Needle 2

Needle is a very small foundation model from Cactus for tool calling and structured extraction on constrained hardware: mobile, wearables, robots, smart home, automotive and microcontrollers. The site page now titled "Needle 3" describes 8–29 MB binaries; a YouTube review of the earlier version described a 45M-parameter model in a 14 MB binary using about 28 MB of RAM and decoding around 500 tokens/s on a Raspberry Pi 5. This entry's name (Needle 2) may refer to that earlier release, and the page content supplied is for the newer Needle 3.

The model uses "intelligence laddering": each layer forms a sub-network of increasing capacity, so developers can pick anywhere from 2 to 20 layers. Cactus reports that fine-tuned sub-networks from 4 layers up (starting at 29M parameters) can match or pass DeepSeek V4 Flash on a DroidCall-style tool-calling task. Responses are JSON objects with function calls, short reasoning, and a calibrated confidence score. Calls below a floor of 0.1 are withheld, and applications can route on the score to act, confirm or refuse. Tools can also have regex triggers that force a call.

Access is through the cactus-compute Python package, with a prebuilt engine under 1 MB for macOS, Linux (x86_64, arm64, armv7, riscv64) and other targets. Fine-tuning is available through LoRA locally or the paid-path Cactus Platform. It is intended for narrow tasks such as device control and extraction, and is not suited to long-horizon reasoning.