Cactus has launched Needle 2, a 14MB open-source agentic large language model (LLM) designed for phones, wearables, smart home devices, and robots. The model features 45 million parameters and runs efficiently in just 28MB of RAM. Needle 2 is built on the Simple Attention Network architecture and compressed using Cactus Quants CQ2-bit technology, enabling fast and lightweight performance on tiny devices, according to cactuscompute.com.
Needle 2 supports tool calling, device control, and structured data extraction with 12 defined tools for tasks such as email extraction, currency conversion, and sentiment classification. It achieves decoding speeds of 500 tokens per second on a Raspberry Pi 5, 400 to 1,500 tokens per second on VR devices like Meta Quest 3S and Apple Vision Pro, and 300 to 700 tokens per second on budget smartphones such as the Samsung A-Series. This performance is competitive with larger models like FunctionGemma 270M and Apple FM, despite being significantly smaller.
The release of Needle 2 highlights advances in deploying capable LLMs on resource-constrained hardware, addressing the growing demand for AI on edge devices. Its compact size and speed make it suitable for real-time applications in smart homes and robotics, where latency and privacy concerns limit cloud dependency. Needle 2's efficiency contrasts with larger models that require more memory and power, marking a notable step in miniaturizing AI for consumer electronics.
Needle 2 is available now on cactuscompute.com, where developers can access the model and experiment with its capabilities. The open-source release invites integration into various tiny devices, potentially expanding AI functionality across a broad range of consumer and industrial applications.