AI Potluck
Back to Gap Map Model components / Multimodal models

Fara

Microsoft
open weights / Overall score: 3.4

Fara is Microsoft's family of computer-use agent models, which look at webpages and act by scrolling, typing and clicking at predicted coordinates without accessibility trees. Fara1.5 comes in 4B, 9B and 27B sizes built on Qwen3.5 and trained on data from the FaraGen1.5 pipeline, and can pause to ask the user before irreversible actions. Fara-7B was the first release.

Openness

3 high confidence
3.0
weights
open(on the Hub, ungated)
data
described(the roughly 2M-sample FaraGen1.5 training mix is broken down by task type
code
partial(agent runner, hosting instructions and the webeval benchmark harness
license
MIT(OSI

Fara's weights and code are MIT, and Microsoft publishes the tools to run and evaluate the agent. The FaraGen1.5 training data is described in detail but not released, and no training code is published.

Adoption

2 high confidence
2.0

Hugging Face downloads over the trailing 30 days, summed across Microsoft's five Fara checkpoints. Use through Microsoft Foundry hosting is not counted.

Capability

4 high confidence
4.0

Fara completes multi-step web tasks by clicking, typing and scrolling on screenshots, and on Online-Mind2Web it scores above GPT-5 in Microsoft's own table. It is level with UI-TARS and, like it, takes no audio or video input.

Verified 2026-09-27