Fara-7B-AIO-GGUF

Fara-7B is Microsoft's first agentic small language model (SLM) with 7 billion parameters, built on Qwen2.5-VL-7B as a multimodal decoder-only architecture specialized for computer use tasks like web automation, shopping, booking travel, restaurant reservations, and account workflows. It processes user goals, browser screenshots, and action history within a 128k token context to generate chain-of-thought reasoning followed by grounded tool calls for actions such as mouse movements, clicks, typing, scrolling, URL visits, and web searches, mimicking human-like desktop interactions without accessibility trees. Trained in just 2.5 days on 64 H100 GPUs using 145k synthetic trajectories from a multi-agent pipeline, it achieves state-of-the-art results in its size class—73.5% on WebVoyager, 38.4% on WebTailBench—outperforming peers like UI-TARS-7B while incorporating safety safeguards to halt at critical points (e.g., purchases, personal info entry) and refuse harmful tasks.[1][2][3][4]

Fara-7B [GGUF]

Quants Usage