Needle is a foundation model built by Cactus Compute for the smallest devices: phones, wearables, smart home hardware, robots, cars, and even microcontrollers. Packaged as a single binary of just 8 to 29 megabytes, it trades general chat ability for something edge devices actually need: reliable tool calling, structured data extraction, embeddings, and speech input, all running offline. The open-source project has earned more than 14,000 stars on GitHub and is under active development.
Needle uses a 2-bit design on the company’s Simple Attention Network architecture, engineered so a very small model beats models ten times its size on mobile tool calls and matches models two to three times larger on extraction tasks. In practice it does four jobs. Given the functions your app exposes, it picks the right tools and fills every argument from natural speech or text. Given a schema, it extracts typed fields from messy text, with a decode grammar that guarantees the output parses. It returns sentence embeddings for local search and routing. And with the companion Whistle speech-to-text model, a voice clip goes in and tool calls come out. Every response carries a calibrated confidence score, which lets apps decide whether to act locally or escalate to the cloud, the hybrid pattern Cactus promotes for production devices. Real products, including the Pebble Index ring, already run Needle on-device.
Needle is open source under the Apache 2.0 license, and its model weights are publicly available through the project’s Hugging Face presence. There is no usage fee for the open models; your costs are your own product’s hardware and engineering.
Needle is for device makers, app developers, and embedded engineers who need dependable AI actions where the cloud is unavailable, too slow, or too private to use. If your product must understand commands and extract data on the device itself, Needle is one of the few models designed exactly for that job.