Microsoft Research has introduced Fara-7B, a groundbreaking Small Language Model (SLM) designed for seamless computer interaction. Unlike traditional chatbots, Fara-7B functions as a Computer Use Agent (CUA), utilizing simulated keyboard and mouse inputs to perform tasks directly on devices like Windows 11 PCs. With only 7 billion parameters, this compact model reduces latency and enhances privacy by processing data locally. Fara-7B visually interacts with screens, predicting actions through user-like visual perception, and boasts a robust training dataset of 145,000 trajectories. On benchmarks like WebVoyager, it achieves a task success rate of 73.5%, outperforming larger models, while averaging only 16 steps per task for efficiency. To ensure safety and control, Fara-7B processes minimal data and logs actions for user oversight, addressing potential risks. Available on Microsoft Foundry and Hugging Face, Fara-7B invites community exploration of automated web task handling, enhancing productivity through innovative CUA technology.
Share
Read more