Microsoft Fara is the official GitHub repository for Fara1.5, Microsoft Research’s family of frontier computer-use agent models. These are native computer-use agents โ models trained to look at a screen, reason about what to do, and operate a browser or desktop by clicking, typing and scrolling. Fara1.5 ships in three sizes (4B, 9B and 27B parameters) built on Qwen3.5 and trained with Microsoft’s FaraGen1.5 data pipeline, with all weights openly available. Note that this is Microsoft’s computer-use agent project, a distinct project from other similarly named products listed on Tooloso. The repository has about 6,200 stars, with the models released in 2026.
Computer-use agents promise to automate real workflows in the interfaces people already use, but they need models that can ground instructions in pixels reliably and recover from mistakes. Fara1.5 was trained for exactly that: each model in the family takes screenshots as input and produces actions, with the three scales trading capability against the cost of self-hosting. Weights are published on Hugging Face and the models can also be hosted through Microsoft Foundry. Alongside the models, Microsoft released supporting benchmarks and datasets such as WebTailBench and CUAVerifierBench to measure how well agents complete and verify web tasks, making this repository a useful hub for anyone evaluating or building on open computer-use models.
Microsoft Fara is free and open source: the repository is MIT-licensed and the model weights are openly downloadable. Your only costs are the compute to run the models or any hosted inference you choose.
Microsoft Fara is for researchers and developers building computer-use agents, teams evaluating open alternatives to proprietary GUI agents, and anyone who needs a self-hostable model that can operate real software interfaces from screenshots.