Cactus Compute released a series of compact Needle 3 models

Company Cactus Compute introduced a line of specialized Needle 3 models designed for automating tasks on devices with limited resources. The weight of the presented neural networks varies from 8 to 29 MB, which allows them to be integrated into smartphones, wearable electronics, and robotic systems.

The technical specifications of Needle 2, the predecessor of the new series, demonstrate high performance at the hardware level. A model with 45 million parameters and a size of 14 MB can generate text at a speed of more than 500 tokens per second on a single-board computer Raspberry Pi 5. On budget smartphones costing less than $200, speed indicators range from 300 to 700 tokens per second.

Developers claim that in terms of efficiency in automation tasks, Needle 3 is comparable to the DeepSeek V4 Flash model. Such a small weight volume is achieved due to the deep optimization of the architecture for specific usage scenarios, where low latency and minimal RAM consumption are critical.

This release is targeted at developers creating autonomous AI agents for local execution. The use of such models eliminates the need to access cloud APIs, ensuring data privacy and system operation without a constant network connection. Detailed technical discussions and community feedback are available on the Hacker News platform.