Needle

Needle is an AI model for developers building device assistants that turn requests into actions. It prioritizes automation over general chat, with smaller model variants for devices that have limited resources.

• Tool calls: It selects functions exposed by an app, fills their arguments and returns ordered calls for requests with multiple actions; unsupported requests can return no calls.
• Structured extraction: It turns invoices, bookings and notifications into typed fields, with output constrained to valid JSON, and supports classification.
• Local search: Text embeddings let apps search notes and documents by meaning, match requests to tools and identify duplicate alerts without sending embeddings off the device.
• Speech input: The companion Whistle model runs speech recognition on the CPU and feeds audio requests into tool calls, with support for English, German, French, Spanish, Italian, Dutch and Polish.
• Confidence controls: Scores help an app decide when to execute an action or ask for confirmation, and the engine can withhold low-confidence calls.

Needle is open source under Apache 2.0. It runs on Apple Silicon macOS, Windows, Linux, Android, iOS, tvOS, watchOS and browsers, including Linux targets for Raspberry Pi and other small devices. Inference can run offline after the model and engine are downloaded. The binary enables telemetry by default; users can disable it.

Developers can fine-tune locally with LoRA adapters or use the hosted Cactus Platform for full-model training, dataset management and evaluation. Platform training runs on Cactus infrastructure. The model's smaller variants let developers trade capacity for a lighter deployment on their target hardware.