Web automation is shifting from code parsing to visual perception. Microsoft Research AI Frontiers has released the weights for Fara1.5-27B, a Computer Use Agent (CUA) specifically designed for web browsers. Unlike traditional automation tools, Fara does not rely on the DOM or accessibility trees; it perceives the interface solely through screenshots, mimicking human visual interaction.

Visual Perception and Structured Action

The core of Fara1.5-27B is its ability to translate visual input into concrete actions. By analyzing the current browser state and cross-referencing it with action history, the model predicts the next step. This results in structured tool calls: clicking specific pixel coordinates, typing text, scrolling, or navigating to new URLs.

The model is a supervised fine-tuned version of Qwen3.5-27B, optimized via a multi-agent pipeline called FaraGen1.5. This system synthesizes complex web tasks and executes trajectories to train the model for autonomous end-to-end completion.

I Put Microsoft's AI Browser to the Test. Here's What Actually Works — https://au.pcmag.com/ai/113938/i-put-microsofts-ai-browser-to-the-test-heres-what-actually-works

Efficiency and the Open-Weight Ecosystem

The release of Fara1.5 aligns with a broader push toward optimizing 27B parameter models. While the FP16 baseline requires roughly 54 GB of VRAM, the community is exploring aggressive quantization. Projects like Bonsai 27B are testing 1-bit representations, reducing the memory footprint by up to 14x to enable local execution on consumer laptops or via WebGPU in the browser.

This openness reflects Microsoft's strategy of integrating open-weight models into Microsoft Foundry, streamlining the deployment of curated Hugging Face models on managed Azure infrastructure.

Toward a Universal Agentic Interface

Fara's approach offers a flexible alternative to deep browser integration. While other players are embedding AI directly into the browser engine, Fara's visual abstraction makes it potentially independent of a website's underlying technical structure.

However, this autonomy introduces security risks. Recent incidents, such as the prompt injection attack on a Grok agent, highlight that an agent's ability to interact with the live web significantly expands the attack surface for AI systems.