Cactus Compute released Needle, a tiny automation foundation model optimized for resource-constrained environments like phones, wearables, and microcontrollers. The entire model is delivered as a single 8-29 MB binary using a Simple Attention Network, trading general chat capabilities for superior performance on specific tasks. It excels at structured extraction and tool invocation, reliably filling arguments from user input without hallucinating calls for unsupported functions.
- Needle fits in 8-29MB, making it viable for microcontrollers and wearables.
- Optimized for tool calls and structured extraction rather than general chat.
- Returns empty lists for unsupported functions instead of hallucinating arguments.
- Outperforms models 10x larger on mobile tool execution tasks.
- Single binary distribution simplifies deployment on edge hardware.
