Fara1.5: Microsoft's Open-Weight Computer Use Agent Models for Browser Automation

What Fara1.5 Solves and Who It Is For

Fara1.5 addresses the cost and opacity of proprietary computer use agents by releasing model weights that researchers and developers can host, fine-tune, and evaluate independently. The three sizes (4B, 9B, 27B) allow teams to choose between memory efficiency and task performance, and the repository includes evaluation harnesses for the WebTailBench and Online-Mind2Web benchmarks so results can be reproduced.

The primary audience is AI researchers evaluating browser automation models, engineers building agent pipelines, and developers exploring the Magentic-UI sandboxed browser environment. It is not intended as a turnkey product for non-technical users.

The Observe-Think-Act Loop and Coordinate Prediction

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 current task state and outputs an action. Actions include mouse and keyboard inputs on directly predicted pixel coordinates, web searches, and context management operations. Notably, the models do not use accessibility trees or separate parsing models: the prediction is made from the raw screenshot.

This approach means that Fara1.5 can interact with any website that renders visually, including those with dynamic JavaScript content that accessibility APIs struggle to parse. The trade-off is that coordinate prediction is sensitive to screen resolution and viewport size changes.

Fara1.5 is also trained with a user simulator on multi-turn rollouts, which means it can ask for missing information, flag ambiguous tasks, and pause for approval before irreversible actions.

FaraGen1.5: The Data Pipeline Behind the Models

The README describes FaraGen1.5 as the scalable data pipeline used to train Fara1.5. It has three modular components: Environments (open-internet tasks on live websites plus six synthetic FaraEnvs covering Mail, Calendar, and others), Solvers (the agents that attempt the tasks), and Verifiers (the judges that score outcomes).

A key verifier is the Universal Verifier, called `MMRubricAgent`, which is also the official judge for the WebTailBench benchmark. The repository includes `webeval/scripts/webtailbench.py` for running WebTailBench evaluations and `webeval/scripts/verify_trajectories.py` for standalone re-scoring of saved trajectories. WebTailBench uses the `test_v2` split; a documented diff between the V1 and V2 splits is hosted at microsoft.github.io/fara/docs/webtailbench_v1_v2_diff.html.

The CUAVerifierBench dataset, released on 2026-04-19, benchmarks the verifiers themselves. It contains 106 Fara-7B Online-Mind2Web trajectories and 154 trajectories from an internal task suite, with per-judge labels side-by-side. The dataset is hosted on Hugging Face at microsoft/CUAVerifierBench with two splits: fara7b_om2w_browserbase and internal.

Installation and Running the CLI

The repository requires Python 3.10 or later. The recommended setup uses a virtual environment:

git clone https://github.com/microsoft/fara.git
cd fara
python3 -m venv .venv
source .venv/bin/activate
pip install -e .
playwright install

After installation, deploy a model from the Microsoft Foundry catalog and create a config file:

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

Then run an interactive session with:

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

For local model serving, the pyproject.toml lists an optional `vllm` dependency: `vllm==0.19.1` with `transformers>=5.2.0`. Windows users should run inside WSL2; the README explicitly recommends this. The Magentic-UI integration requires a separate setup following the Magentic-UI repository instructions.

Benchmarks Published in the README

The README reports specific benchmark results for the Fara1.5 family. Fara1.5-9B reaches 63.4% on Online-Mind2Web and 86.6% on WebVoyager. Fara1.5-27B achieves 72.3% on Online-Mind2Web and, according to the README, outperforms proprietary systems including OpenAI Operator and Gemini 2.5 Computer Use on that benchmark.

These numbers should be read as the figures the Microsoft team reported at release time, July 2026. Independent replication requires setting up the WebTailBench evaluation pipeline using the `test_v2` split, as the README notes that the `test_v1` split has a documented diff from `test_v2`.

A limitation worth noting is that the benchmarks cover browser automation on public websites. Performance on internal enterprise applications, sites with heavy JavaScript frameworks, or multi-step tasks requiring external API calls is not documented in the repository.

Limitations and When to Look Elsewhere

Fara1.5 is explicitly labeled a research preview. The repository includes a `TRANSPARENCY_NOTE.md` and a `SECURITY.md`, and the README lists safety measures including task refusal based on responsible AI policies and auditable action logging through Magentic-UI.

Coordinate-based action prediction creates a specific failure mode: any change in viewport size, screen DPI, or dynamic layout reflow can shift element positions and cause misclicks. Applications with animations or heavy CSS transitions that alter element coordinates between frames are particularly risky.

The vllm serving stack pins to `vllm==0.19.1`. The pyproject.toml comment notes that `nvidia-cutlass-dsl` is capped at 4.5.2 because version 4.6.0.dev0 removed an alias that a transitive vLLM dependency uses, crashing vLLM at startup. Teams on newer CUDA toolchains will need to verify compatibility before upgrading the serving stack.

OpenAI Operator is a comparable commercial alternative. It provides a managed API without requiring local model hosting or a Microsoft Azure endpoint, which lowers the operational burden. The trade-off is that model weights are not available for fine-tuning or offline evaluation.

The repository also does not include Docker images, Helm charts, or other deployment artifacts. Teams that want to serve Fara1.5 in a production container environment must build their own container setup. The `pyproject.toml` classifies the package as Development Status Alpha, which reflects the research preview designation accurately and should inform production deployment decisions.

License and Repository Maintenance

The repository is MIT licensed and is not archived. The last push was on 2026-09-22. Model weights are hosted separately on Hugging Face under the repositories listed in the README; the repository itself contains only code and configuration files.

There are no GitHub releases. The pyproject.toml classifies the package as Development Status Alpha, which accurately reflects the research preview designation in the README. The repository includes a `CODE_OF_CONDUCT.md` and a `SECURITY.md`, consistent with standard Microsoft open-source practices. The top-level directories include `src/` for the main Python package code, `tests/` for the test suite, `webeval/` for the benchmark evaluation harness, and `docs/` for supplemental documentation.

Editorial conclusion

Fara1.5 suits ML researchers and engineers who want to run or fine-tune open-weight computer use agents locally or via Microsoft Foundry, and who are comfortable with Python 3.10+ and Playwright. It is not a ready-made product for end users: there is no packaged application and the README notes it remains a research preview with responsible AI policies still evolving. Before deploying, confirm your hardware can serve at least the 9B variant; the 4B model fits on modest GPU memory but the README does not specify VRAM requirements per model size, so you must test your setup with the vllm optional dependency against your target hardware.

Frequently asked questions

Official sources

1. Issues
2. License: MIT
3. microsoft/fara on GitHub
4. Project website
5. README