Fara-7B

Fara-7B is Microsoft’s first agentic small language model designed specifically for computer use. The 7-billion-parameter model is built on Qwen2.5-VL, perceives the screen through screenshots, and predicts coordinate-grounded actions directly. It reaches state-of-the-art performance within its size class on WebVoyager, Online-Mind2Web, DeepShop, and WebTailBench, competing with substantially larger agentic systems. Training relies on large-scale synthetic action trajectories generated by multi-agent pipelines, and the model includes Critical Point recognition to pause at sensitive operations — checkout, payment, authentication — for explicit user approval. Pretraining used 64 H100 GPUs over 2.5 days, with a 128K context window.

Fara-7B demonstrates that compact, on-device-friendly models can match or exceed much larger systems when trained on high-quality, action-grounded data. Direct coordinate prediction sidesteps the brittleness of abstract action languages, and the synthetic trajectory pipeline removes the human-annotation bottleneck that has historically slowed agentic research. Built-in safety checkpoints give users a meaningful seam of control over the model’s most consequential actions. Together these choices make Fara-7B a credible foundation for everyday computer-use agents that can run efficiently on consumer hardware.