7B parameters, 2.5 days of H100 training, and MIT open source—Microsoft’s Fara turns “AI operating computers” from a demo into deployable infrastructure. The key is that it also does safety alignment to refuse rule-violating tasks, which is what a real agent should look like.

Microsoft releases Fara-7B, a multimodal intelligent agent with 7 billion parameters, designed specifically for computer usage scenarios. It can process screenshots and text simultaneously, directly predicting parameterized thought chains and operational actions, built on Qwen 2.5-VL, with a 128k context window, trained on 64 H100s for 2.5 days, released under MIT license. It perceives browser input through screenshots, combining reasoning and historical state prediction to determine the next actions and parameters such as coordinates, relying on large-scale fully synthetic data. It has the capability to plan and execute advanced tasks, employing robust post-training safety alignment, able to refuse inappropriate tasks and pause at critical points. It can be deployed and interacted with via GitHub, vllm, and fara-cli, used for automating web page tasks.