Cactus Releases Needle 3: A Quantized 8–29MB Model Matching DeepSeek on Embedded Devices

Cactus Compute has released Needle 3, an automation foundation model compressed to just 8–29 MB that achieves performance parity with models orders of magnitude larger when fine-tuned for on-device task automation. The model, built as a Laddered Simple Attention Network with a novel Monarch Hadamard MLP architecture and CQ2 quantization (approximately 2.125 bits per weight), represents a significant step toward deploying capable AI agents on phones, wearables, smart home systems, and microcontrollers without cloud dependency.