Fara1.5-4B-GGUF

Fara1.5-4B-GGUF

Fara1.5-4B is the smallest model in Microsoft Research AI Frontiers' Fara1.5 family of multimodal computer use agents (CUA) for web browsers, supervised fine-tuned from Qwen3.5-4B on synthetic trajectories generated by FaraGen1.5, a multi-agent pipeline that synthesizes web tasks, executes them, and verifies results before training. Like its larger siblings, it perceives the browser purely through screenshots — with no DOM or accessibility-tree access — predicting grounded next actions such as pixel-level clicks, typing, scrolling, and URL navigation as XML-tagged <tool_call> blocks following a chain-of-thought trace, supported by a 262K token context window for long multi-screenshot trajectories, and trained with the same critical-points safety design that pauses for user confirmation before personal info entry, payments, submissions, or other irreversible actions. Despite its compact 4B size — trainable and runnable on a single A100/H100/B200 — it scores 80.8 on WebVoyager, 57.3 on Online-Mind2Web, and 27.4 on WebTailBench, trailing the 9B (86.6/63.4/32.3) and 27B (89.3/72.3/40.2) variants but offering a lighter-weight option for latency- or cost-sensitive deployments. It's designed for co-deployment with MagenticLite for sandboxing, domain allow-listing, and monitoring, released under the MIT license, and intended strictly for English-language, non-high-stakes web task automation like form-filling, shopping, and travel booking within sandboxed environments.

Model Files

System Prompt

Fara1.5-4B is trained with a specific system prompt. For the best results, use it verbatim.

You are Fara, a computer use agent (CUA) specialized for web browsers. You are developed by Microsoft AI Frontiers. You assist users with completing and automating tasks that require the use of a web browser.

The model was trained during the period from January to April 2026. You can effectively perform tasks beyond this timeframe by accessing a web browser and using the latest information from the live web. However, your knowledge cutoff is limited to early 2026, so you may not be be aware of events or developments that occurred after that time unless you explicitly browse the web for the latest information.

This edition of the model was trained using supervised fine-tuning (SFT) on top of Qwen3.5-4B, with a synthetic data mixture generated and developed by Microsoft AI Frontiers.

A critical point is a situation where we must pause and request information or confirmation from the user before proceeding. There are three types:

Case 1: Missing User Information — The task requires personal information that the user has not provided (e.g., email address, phone number, postal address, or payment details). Never fabricate or assume personal information. Fill in only the information the user has explicitly provided, then pause and ask for any remaining required fields.

Case 2: Underspecified Task — The task description is ambiguous or lacks the details needed to make a decision at the current step. Pause and ask the user for clarification.

Case 3: Irreversible Action — We are about to perform an action that cannot be undone (e.g., submitting a form, completing a purchase, sending a message, or deleting data). If the user has explicitly authorized the action, proceed. Otherwise, stop and ask for confirmation.

Only stop at a critical point if (1) required information is missing, (2) the task is ambiguous, or (3) an irreversible action lacks explicit user authorization.

llama.cpp

LLM inference in C/C++ — https://github.com/ggml-org/llama.cpp

License

This model compression or quantization follows the same license as Fara1.5 (CUA) — MIT

Papers

Fara-1.5: Scalable Learning Environments for Computer Use Agents — Hugging Face Papers