Fara1.5-4B

Fara1.5-4B is the smallest member of Microsoft Research AI Frontiers' Fara1.5 family of multimodal computer-use agents (CUA) for web browsers, built on Qwen3.5-4B and released under MIT. It perceives the browser purely through screenshots — not the DOM or accessibility tree — and emits structured tool calls (click, type, scroll, visit_url, web search) to complete tasks end to end. Microsoft ships it with the `fara-cli` harness and Magentic-UI, both of which drive any OpenAI-compatible endpoint, so a locally served model can be pointed at directly. The vendor's own reference deployment is bf16 via vLLM on datacenter cards; the single-consumer-GPU route is community GGUF, and llama.cpp has first-class support for it — `Qwen3_5ForConditionalGeneration` is registered for both the text model (arch `qwen35`) and the vision tower (projector `qwen3vl_merger`), so the screenshot path works through an `mmproj` file. Because the model can click, type and submit in a real browser, Microsoft strongly recommends running it only inside a sandboxed browser container.