fara

Updates

• 2026-07-22 — Fara1.5 released! A family of native computer use agents at three scales (Fara1.5-4B, Fara1.5-9B, Fara1.5-27B) built on Qwen3.5 and trained on data from the FaraGen1.5 pipeline. All three models have weights available on Hugging Face and can be hosted through Microsoft Foundry (Fara1.5-4B, Fara1.5-9B, Fara1.5-27B). Fara1.5-9B reaches 63.4% on Online-Mind2Web and 86.6% on WebVoyager, a new state of the art for its size class; Fara1.5-27B achieves 72.3% on Online-Mind2Web, outperforming much larger proprietary systems. Read the paper and the blog post.
• 2026-05-12 — WebTailBench is now a first-class benchmark in this repo: the loader auto-downloads tasks and rubrics from microsoft/WebTailBench (use the refreshed test_v2 split; a V1↔V2 diff is hosted here), the Universal Verifier (MMRubricAgent) is the official judge, and the reproducibility CLI lives in webeval/scripts/webtailbench.py (stand-alone re-scoring via webeval/scripts/verify_trajectories.py).
• 2026-04-19 — Released CUAVerifierBench, a human-annotated benchmark for evaluating CUA verifiers (i.e. judges that score agent trajectories). Two splits — fara7b_om2w_browserbase (106 Fara-7B Online-Mind2Web/Browserbase trajectories, ~2 reviewers each) and internal (154 trajectories from a heldout aurora-v2 task suite) — with per-judge UV-blind / UV-informed labels, Universal Verifier outputs, and legacy verifier outputs side-by-side. The build script that produced the dataset lives alongside the data on Hugging Face.

Overview

Fara1.5 is a family of native Computer Use Agents (CUAs) at three scales — Fara1.5-4B, Fara1.5-9B, and Fara1.5-27B — built on Qwen3.5 and trained with supervised finetuning on data from FaraGen1.5, our scalable data pipeline of environments, solvers, and verifiers. Each model sets a new state of the art for its size class on browser-use benchmarks, and Fara1.5-27B outperforms much larger proprietary systems such as OpenAI Operator and Gemini 2.5 Computer Use on Online-Mind2Web.

Fara1.5 models operate through an observe-think-act loop: given a screenshot of the browser and the conversation history, the model reasons about the state of the task and outputs an action — mouse and keyboard inputs on directly predicted coordinates, web searches, or context management operations — with no accessibility trees or separate parsing models.

All three models are available on Microsoft Foundry: Fara1.5-4B, Fara1.5-9B, and Fara1.5-27B.

Try Fara1.5-9B as follows (see Installation for detailed instructions):

# 1. Clone repository
git clone https://github.com/microsoft/fara.git
cd fara

# 2. Setup environment
python3 -m venv .venv 
source .venv/bin/activate
pip install -e .
playwright install

Deploy Fara1.5-9B from the Microsoft Foundry catalog and put your endpoint in a config JSON (e.g. azure_foundry_config.json):

{
    "model": "Fara1.5-9B",
    "base_url": "https://your-endpoint.inference.ml.azure.com/",
    "api_key": "YOUR_API_KEY_HERE"
}

Then you can iteratively query it with:

fara-cli --task "whats the weather in new york now" --endpoint_config azure_foundry_config.json

To try Fara inside Magentic-UI — a sandboxed browser environment with auditable action logging and user prompts at critical points — follow the instructions in the Magentic-UI repo. You will need a model endpoint as before, but instead of fara-cli you can use Magentic-UI which has a nice UI (see video demos below).

Note: If you're using Windows, we highly recommend using WSL2 (Windows Subsystem for Linux). Please see the Windows instructions in the Installation section.

What Makes Fara1.5 Unique

Unlike traditional chat models that generate text-based responses, Fara1.5 leverages computer interfaces—mouse and keyboard—to perform multi-step tasks on behalf of users. The models:

• Operate visually by perceiving webpages and taking actions like scrolling, typing, and clicking on directly predicted coordinates without accessibility trees or separate parsing models
• Enable efficient deployment thanks to their compact sizes (4B/9B/27B), resulting in reduced latency and improved privacy
• Interact with users: trained with a user simulator on multi-turn rollouts, Fara1.5 can ask for missing information, flag ambiguous tasks, and pause for approval before irreversible actions
• Remain a research preview: safety measures include refusing harmful tasks based on responsible AI policies, prompting users at critical interaction points, and auditable action logging through Magentic-UI's sandboxed browser environment

FaraGen1.5: Scalable Learning Environments

Collecting computer use data from human demonstrations is expensive and slow. FaraGen1.5 is a scalable data pipeline composed of three modular components:

• Environments — open-internet tasks on live websites plus six synthetic FaraEnvs (Mail, Calendar, Stream, ML, Stay, Scheduler): functional UI clones that faithfully simulate domains gated by authentication or requiring irreversible actions, and unlock execution-based, ground-truth verification
• Solvers — a solver harness that can be powered by multiple models, including strong frontier models such as GPT-5.4, paired with a user simulator to enable multi-turn rollouts
• Verifiers — three complementary filters covering task correctness (the Universal Verifier LLM-judge), efficiency (scoring redundant actions), and critical-point adherence / user interaction (missing information, ambiguous tasks, unapproved irreversible actions)

The resulting training mix contains roughly 2M samples: ~60% web trajectories, 12.8% synthetic environments, 12.5% form filling, 8.8% grounding, 4.9% VQA, and 0.8% GUI drag tasks.

Key Capabilities

Fara1.5 can automate everyday web tasks including:

• Searching for information and summarizing results
• Filling out forms and managing accounts
• Booking travel, movie tickets, and restaurant reservations
• Shopping and comparing prices across retailers
• Finding job postings and real estate listings

Performance Highlights

Each Fara1.5 model sets a new state of the art for its size class. Fara1.5-9B improves over Fara-7B by +29.3 points on Online-Mind2Web, +13.1 points on WebVoyager, and +8.2 points on WebTailBench outcome success:

Table: Task success rate (%) on WebVoyager, Online-Mind2Web, and WebTailBench v1.5. For WebTailBench v1.5 we report both Process Success (correct intermediate steps) and Outcome Success (final task state correct). All Fara1.5 and Fara-7B numbers are averaged over three independent runs. The GPT-5.4-based FaraGen1.5 solver is an upper-bound reference for the SFT-based distillation. † denotes numbers sourced from the model's official release or leaderboard rather than re-run by us.

WebTailBench: Evaluating Real-World Web Tasks

WebTailBench is our evaluation benchmark focusing on 11 real-world task types that are underrepresented or missing in existing benchmarks. The benchmark includes 609 tasks across diverse categories, with the first 8 segments testing single skills or objectives (usually on a single website), and the remaining 3 evaluating more difficult multi-step or cross-site tasks (shopping lists, comparison shopping, and compositional tasks). Tasks and precomputed rubrics are on Hugging Face (use the refreshed test_v2 split), the Universal Verifier (MMRubricAgent) is the official judge, and the reproducibility CLI lives in webeval/scripts/webtailbench.py.

CUAVerifierBench: Evaluating the Verifiers Themselves

While WebTailBench measures agents, CUAVerifierBench measures the judges that score those agents. Each row pairs a Fara agent trajectory (instruction, screenshots, web_surfer log, final answer) with one human reviewer's verdict, plus the verdicts produced by the Universal Verifier (MMRubricAgent) and several legacy verifiers — so researchers can compute verifier–human agreement (Cohen's κ, accuracy, F1) on a fixed corpus and iterate on new judge prompts / architectures against a frozen ground-truth set.

The dataset is exposed as two Hugging Face configs joinable on task_id:

Two splits ship today:

Reviewer identities are anonymized as Judge1 … JudgeN using a single shared map across both splits. The build script that produced the dataset (with full schema + provenance) ships alongside the data on Hugging Face at [microsoft/CUAVerifierBench](https://huggingface.co/datasets/microsoft/CUAVerifierBen