Needle 3

Needle 3 is a foundation model designed for tiny devices, including mobile phones, wearables, robots, smart home systems, automotive systems, and microcontrollers. It solves the problem of relying on large, cloud-based AI by offering a single 8-29 MB binary that runs locally on the device. The model beats models ten times its size on mobile tool calls and matches models two to three times larger on extraction tasks.

It handles tool calls, structured extraction, and text embedding, allowing devices to execute actions offline without requiring a hub or network connection. The system parses messy text into structured fields and generates vectors for local semantic search, enabling the device to understand requests and perform tasks independently.