Autonomous web agents often trade off fidelity to real UIs versus reliance on internal page state. Fara1.5-27B intentionally adopts a human-like, vision-first approach: it observes browser screenshots, maintains textual reasoning and history, then emits coordinate-grounded tool calls to progress multi-step tasks. This lets it operate where DOM or accessibility trees are unavailable, but it also makes sandboxing and human oversight essential because visual ambiguity and irreversible actions can cause costly mistakes.
Key Capabilities
- Vision-only perception with coordinate grounding: predicts pixel-level click/drag targets and typed text directly from screenshots, removing the need for separate grounding modules.
- End-to-end web task completion: trained to sequence actions toward goals (form filling, bookings, shopping, job applications) rather than isolated clicks; emits XML-tagged tool_call blocks compatible with a computer_use schema.
- Large context and model scale: 27B parameters and a 262,144-token context enables multi-screenshot trajectories and long reasoning histories for multi-step flows.
- Safety-by-design: trained to pause at three critical-point classes (missing user info, underspecified decisions, irreversible actions) and to avoid entering credentials or completing purchases without explicit authorization.
- Deployment-ready integration: supervised fine-tune on FaraGen1.5 trajectories and co-designed with MagenticLite, which provides sandboxing, allow-lists, watch-mode, and immediate pause controls for safer runs.
Who it's for and tradeoffs
Great fit if you need an agent that can interact with arbitrary web UIs where DOM access is not possible, want pixel-accurate action prediction, and can enforce sandboxing and human-in-the-loop review. It is useful for prototyping automated browser workflows, research into multimodal agents, or as a grounding/action model inside larger agent stacks.
Look elsewhere if you require robust multi-language support (training is English-only), production deployments without strict sandboxing or allow-lists, high-stakes decision-making (legal/medical/financial), or absolute determinism across long trajectories—visual misrenderings and run-to-run variance remain failure modes. The model can also hallucinate page state and should not be given unrestricted access to sensitive accounts or host resources.
Additional notes: released under the MIT License (May 21, 2026); recommended serving patterns include vLLM and MagenticLite. Key training details: SFT on Qwen3.5-27B using synthetic multi-agent trajectories (FaraGen1.5), ~1B text tokens and up to ~1B images, trained Jan–Apr 2026.
