Summary
Fahd Mirza installs and live-tests Microsoft’s newly released Fara 1.5 27B, a vision-only browser automation model that controls a web browser purely by observing screenshots and predicting pixel coordinates for clicks, typing, and scrolling — no DOM access, no accessibility tree. The 27B parameter version handles a 262k context window and is benchmarked as the top open-weight web agent at its size class.
The installation runs on an Nvidia H100 with 80GB VRAM (the model consumes just over 76GB), served via vLLM with Playwright for browser control. Mirza walks through the complete setup from conda environment creation to model serving, then demonstrates two live tasks: looking up Wikipedia’s total page count and navigating YouTube to find the most popular video on his own channel. Both tasks complete successfully, with a full audit log of screenshots and action traces available after each run.
On benchmarks, the 27B version scores 72.3 on Online Mind2Web, outperforming Gemini 2.5 Computer Use, Operator, and Navigator — all closed proprietary systems — making it a notable result for the open-weight ecosystem. Mirza also contextualizes the model’s architecture: each inference step ingests the three most recent browser screenshots plus full conversation history, issues a single atomic action, then repeats, enabling long-horizon task completion without an external scaffold. He recommends the 27B variant for production use, noting the 4B version is unlikely to meet real-world task demands.
📺 Source: Fahd Mirza · Published July 29, 2026
🏷️ Format: Hands On Build
