microsoft/Fara1.5-27B

VISIONConcurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 17, 2026License:mitArchitecture:Transformer0.1K Open Weights Featherless Exclusive Cold

Fara1.5-27B is a 27 billion parameter multimodal computer use agent (CUA) developed by Microsoft Research AI Frontiers, designed for web browsers. It operates vision-only, perceiving the browser through screenshots and executing tasks end-to-end by emitting structured tool calls like clicks, types, and scrolls. Fine-tuned from Qwen3.5-27B, it excels at automating complex web tasks such as form filling, booking, and shopping, with a notable 262K token context length.

Loading preview...

Overview

Fara1.5-27B is a 27 billion parameter multimodal computer use agent (CUA) from Microsoft Research AI Frontiers, specifically engineered for web browser automation. It functions by interpreting screenshots (vision-only perception) and generating structured tool calls (e.g., click, type, scroll, visit URL) to complete tasks. This model is supervised fine-tuned from Qwen3.5-27B using data generated by FaraGen1.5, a multi-agent pipeline that synthesizes, executes, and verifies web task trajectories.

Key Capabilities

  • End-to-end web task completion: Capable of complex sequences like filling forms, booking reservations, and managing shopping carts.
  • Vision-only perception: Operates solely on screenshots, mimicking human interaction without requiring DOM access.
  • Coordinate-grounded actions: Predicts precise pixel-level targets for clicks and drags.
  • Safety by design: Trained to pause and seek user confirmation before critical actions such as entering personal information, making payments, or submitting irreversible forms.
  • Extended context length: Features a 262,144 token context, supporting multi-screenshot trajectories and full action history.
  • On-device viability: The 27B parameter size allows it to run on a single A100/H100/B200 GPU.

Good For

  • Automating repetitive web tasks like form filling, online shopping, travel booking, and information gathering.
  • Serving as a grounding model for other agents requiring pixel-accurate action prediction.
  • Deployment within sandboxed environments like MagenticLite for enhanced security and monitoring.