Fara-7B: An Efficient Agentic Model for Computer Use (www.microsoft.com)

Microsoft has unveiled Fara-7B, a groundbreaking ultra-compact agentic model designed for efficient computer use. Unlike typical chat-based models, Fara-7B employs a Computer Use Agent (CUA) architecture that enables it to interact directly with user interfaces through keyboard and mouse actions, completing tasks such as web searches, form filling, and travel booking. With only 7 billion parameters, it achieves state-of-the-art performance, rivaling larger models while maintaining the advantages of reduced latency and enhanced privacy by executing tasks locally on devices. What sets Fara-7B apart is its innovative training approach, utilizing a synthetic data generation pipeline that simulates multi-step web tasks sourced from real user interactions. This allows it to perform effectively across diverse benchmarks, including its own WebTailBench for underrepresented task types. The model, which processes only the visible elements of web pages, promotes transparency and user safety by being auditable, encouraging responsible usage in sandboxed environments. Fara-7B’s release empowers developers and researchers to explore new applications for automating routine online activities, potentially transforming how we interact with information and services on the web while addressing key challenges in agent safety and reliability.