Microsoft Publishes Fara 1.5 Browser-Agent Weights Under MIT

Microsoft released 4B, 9B, and 27B Fara 1.5 computer-use agent weights. The models observe browser screens and produce clicks, typing, searches, or requests for user input.

What is newly available

Microsoft Research updated its May announcement on July 22, 2026 to confirm that Fara 1.5 weights in 4B, 9B, and 27B sizes are publicly available on Hugging Face. The repositories are accessible without sign-in and list the MIT license. Each model observes browser screenshots and conversation history, reasons, and emits a single action such as a click, keystroke, or search.

Reported performance

Microsoft reports Online-Mind2Web task success rates of 57.3% for 4B, 63.4% for 9B, and 72.3% for 27B. On WebVoyager it reports 80.8%, 86.6%, and 88.6%. Fara 1.5 results average three runs, while comparisons use the official leaderboard as of May 2026 and Microsoft's automated evaluation setup.

Why it matters

Three sizes give builders options for self-hosted browser agents under different hardware and accuracy constraints. There is no fee for the weights, but GPU, browser infrastructure, proxies, CAPTCHA handling, and maintenance still cost money. Open weights also do not mean a finished free consumer service.

Safety boundary

Fara 1.5 was trained to ask when information is missing, instructions are ambiguous, or an irreversible action needs approval. That behavior is not a guarantee. Logins, payments, email sending, form submission, and deletion need separate pre-execution approval, allowlisted domains and tools, sandboxing, audit logs, and an emergency stop.

Official source