Cactus Compute Releases Needle 2: A 45M-Parameter Tool-Calling Model for On-Device Deployment

Cactus Compute has released Needle 2, a 45-million-parameter open-source model designed for tool calling, device interaction, and structured data extraction on hardware with minimal computational resources. The model ships as a 14 MB binary and runs a full session in approximately 28 MB of RAM, achieving 500 tokens per second on a Raspberry Pi 5 and 300-700 tokens per second on phones under $200. The compact architecture uses a simplified attention mechanism with Hadamard MLP layers and engram key-value memory, requiring only 70 MFLOPs per token. Weights are available in multiple formats including native, GGUF, ONNX, and MLX, compatible with llama.cpp, MLX, vLLM, and other inference engines.