Microsoft Releases Fara1.5-27B, a 27B Vision-Only Web Browsing Agent with 262K Context

Microsoft Research AI Frontiers has released Fara1.5-27B, a 27-billion-parameter multimodal model built specifically to act as a computer use agent (CUA) inside web browsers. The model was published on Hugging Face on 21 May 2026 under the MIT license.

Fara1.5-27B perceives the browser exclusively through screenshots — it does not read the DOM or accessibility tree. Given a user goal, the current screenshot, and a history of prior actions, it outputs a chain-of-thought block followed by an XML-tagged tool call (click, type, scroll, visit_url, web_search, and others) with grounded pixel coordinates for actions like clicks and drags. According to Microsoft, this vision-only approach mirrors the input modality available to a human user and removes the need for a separate grounding model.

Architecture and training

Fara1.5-27B is a supervised fine-tune of Qwen3.5-27B, trained on synthetic data generated by Microsoft's internal pipeline, FaraGen1.5, which synthesizes web tasks, executes trajectories, and verifies results before they're used for training. Training ran from January 2026 to April 2026 on 64 NVIDIA B200 GPUs over six days. The model supports a 262,144-token context window, which Microsoft says is sufficient for multi-screenshot trajectories with full action history — though the reference agent loop retains only the three most recent screenshots per turn.

Pricing is not disclosed since the model is released as open weights; deployment cost depends on self-hosted inference infrastructure. Microsoft recommends sharding the 27B model across at least two GPUs in bf16, with A6000, A100, H100, and B200 cards tested. Serving is supported via vLLM (version 0.19.1+) with a max model length of 262,144 tokens.

Safety design: critical points

Microsoft built Fara1.5-27B to pause and request user confirmation at three categories of "critical points": missing personal information (never fabricating fields like email or payment details), underspecified tasks that require clarification, and irreversible actions such as submitting forms, completing purchases, sending messages, or deleting data. The model will proceed past these points only with explicit prior authorization from the user.

The company states the model is co-designed with MagenticLite, a companion deployment framework providing Docker-based sandboxing, domain allow-lists, real-time watch-mode monitoring, and an immediate pause control. Microsoft explicitly warns that developers integrating Fara1.5-27B directly, without MagenticLite's controls, are responsible for implementing equivalent safeguards, and cautions against running the model with unrestricted browser access on machines containing sensitive data.

No third-party benchmark scores were included in the release materials, so performance claims relative to other computer-use agents remain unverified.

What this means

Fara1.5-27B is Microsoft's entry into the increasingly crowded field of browser-operating agents, competing conceptually with efforts like Anthropic's computer use tooling and OpenAI's agentic browsing features. The 27B parameter count and MIT license make it notably more accessible for self-hosting than closed frontier agents, and the vision-only design sidesteps dependency on site-specific DOM access — a common failure point for scraping-based agents. The explicit critical-points safety framework, requiring pauses before payments, submissions, and irreversible actions, addresses a real liability concern for autonomous browser agents, though its effectiveness in practice depends on how reliably the model recognizes those situations across the long tail of real websites. Because no independent benchmark data accompanies this release, claims about task completion reliability should be treated as unverified until third-party evaluation appears.