Choose theme
AI Resources
Fara1.5-27B
Fara1.5-27B is a Microsoft Research multimodal computer-use model for web browsers. It reads screenshots and emits structured actions such as clicks, typing, scrolling, navigation, and web search.
The official model card lists a 27B Qwen3.5 base, a 262,144-token context window, public weights, a vLLM serving path, and MagenticLite as the recommended deployment. Microsoft also publishes the Fara agent harness. Its article identifies Fara1.5-9B as the hosted Microsoft Foundry option; that is a different checkpoint from this 27B model. Use this as a first read, not a recommendation. Open the original project before trusting details like terms, limits, privacy, cost, setup, or safety.
What it is
A screenshot-driven browser agent model
Fara1.5-27B sees browser screenshots rather than the DOM or accessibility tree, then predicts pixel coordinates and structured computer-use actions for the next step in a task.
Why it stands out
Browser actions with human checkpoints
Its system prompt and training define critical points where the model should ask for help when required personal information is missing, a choice is ambiguous, or an action could be irreversible, such as a purchase, submission, sign-in, or message.
Availability
27B weights and harness; a separate hosted 9B option
Readers can download this 27B model from Hugging Face, use the Microsoft Fara harness and self-host with vLLM. MagenticLite adds a sandboxed browser, domain controls, action logs and a pause control. The Microsoft article's Foundry option is Fara1.5-9B, rather than a hosted copy of this 27B checkpoint.
Why it matters
What makes it useful
Fara1.5-27B gives builders a downloadable model for multi-step browser work rather than a closed assistant alone. Its screenshot-only design, action schema, reference harness, and stated hardware needs make the real tradeoff clearer: the model can work across visual interfaces, but it still needs a controlled browser loop, enough GPU memory, and close human review.
What to know
Where it fits
This supplies the model layer of browser automation. Microsoft asks users to run it with the Fara CLI or MagenticLite; MagenticLite is the recommended deployment. This 27B self-hosting path needs the documented GPU capacity and browser controls. The article separately lists Fara1.5-9B on Foundry for a hosted family option.
Notable points
What stands out
Microsoft reports strong WebVoyager, Online-Mind2Web, and WebTailBench results for Fara1.5-27B. Those benchmark results are project-reported and multi-step runs still vary; the model card says early misclicks can compound, page state can be hallucinated, and deceptive page content or prompt injection can redirect behavior.
Before using
What to review
For self-hosting, check the current model-card requirements for PyTorch, Transformers, vLLM, BF16 GPU memory, context length, and the recommended multi-GPU sharding path.
Microsoft documents a specific system prompt and tool schema, with 1440 x 900 as the usual training resolution for coordinate grounding.
Microsoft recommends sandboxed browsing, restricted domains, action monitoring and a fast stop control. Its model card excludes unsandboxed access to sensitive accounts or files from intended use.
Microsoft describes critical points for missing personal information, ambiguous tasks and irreversible actions without user authorization. The same card documents prompt-injection and hallucination risks; this training is not a guarantee that every checkpoint will work.
The model card lists English-only training, high-stakes decisions, unsandboxed sensitive access, and production use without additional testing and safeguards as outside its intended scope.
Keep host files, environment secrets, credentials and real accounts outside the browser's reach unless the task specifically requires and authorizes access. A model checkpoint does not create that access boundary.
Require explicit human authorization before forms are submitted, purchases made, messages sent, sign-ins completed or personal information supplied. Review the proposed action and keep a stop control available rather than relying on the model to ask at every critical point.
Reader fit
Who may find it relevant
Builders comparing public-weight computer-use models for browser automation.
Teams evaluating the 27B self-hosted vLLM path, or comparing it with the distinct hosted Fara1.5-9B family option.
Researchers testing screenshot-only grounding, long browser trajectories, or human confirmation points.
Less relevant for readers who need a lightweight local model, desktop-wide control, or a ready-made consumer assistant.
Editorial note
Why LifeHubber lists it
A browser task that compares information across pages needs facts from an earlier page after navigation. Microsoft's Fara loop keeps the text conversation and provides actions for retaining facts or asking the user, alongside a limited set of recent screenshots. That makes cross-page fact handling a concrete part of its browser workflow, while the retained text still needs checking against the source pages. Reproducing a long browser run means keeping its action history as well as its screenshot inputs. Microsoft's sample loop retains only the three most recent screenshots while tracking earlier reasoning and actions as text. An earlier page can therefore remain in the history without its image still being part of the model's visual input.
Source links
Source materials
Reader note
Before relying on this entry
LifeHubber lists entries to help readers inspect AI projects, not to endorse them or prove they are safe, suitable, accurate, maintained, or right for a specific use. We do not verify every entry in depth. Before relying on anything listed, review the original materials, terms, privacy practices, limits, and risks that matter for your situation.
What to explore next
Compare the model with the wider computer-use stack
Continue with LifeHubber's broader computer-use map, another public model family, or infrastructure for controlled desktop and browser agents.
More in AI Agents
Keep browsing this category
Explore more AI agent projects.
Paperclip
paperclipai/paperclip
A self-hosted server and dashboard for coordinating agent teams through companies, goals, roles, issues, heartbeats, budgets, approvals, and persistent activity records.
Universal Modder
rehan-remade/universal-modder
Agent skills and a Python CLI for PC game modding, with engine research, generated assets, in-game checks, and shared field notes. Setup and compatibility depend on the game and workflow; fal generation requires a separate API key.
Sentrux
sentrux/sentrux
A Rust codebase sensor that maps dependencies, scores five structural signals, enforces architecture rules, and exposes scans and before-and-after session checks to coding agents through MCP.
For project maintainers
Listed here? You can use the badge.
If you maintain a project with a current LifeHubber listing, you may add the optional “Listed on LifeHubber AI Resources” badge to its README, docs, or website. No introduction or permission request is needed.
