Give it a goal. A 2B model running in this tab reads the page, picks an element, clicks or types,
and keeps going until it has the answer. Every step re-reads the page, so it only works when the model is fast.
●about:blank
loading page…
Engine
Not loaded.
Goal
Answer
Steps
–
Model time
–
Per step
–
Trace
The agent's thoughts and actions appear here.
1 · Observe
The page in the frame is read through the DOM: numbered links, inputs and buttons, plus a window of its text.
2 · Decide
The local model gets the goal, the page and its previous steps (~1–2k tokens) and answers with one action.
3 · Act
The action runs on the DOM: click, type and submit, search the text, scroll, go back, or answer.
Measured: "When was the Eiffel Tower completed?" (2 steps, ~600 + ~1,400 prompt tokens)
webyIQ · Qwen3.5-2B · RTX 50800.55 s
webyIQ · MiniCPM5-2B · RTX 50800.76 s
transformers.js · MiniCPM5-2B · RTX 50801.9 s
webyIQ · Qwen3.5-2B · Radeon 2-CU iGPU20 s
transformers.js · Qwen3.5-2B · RTX 5080234 s
Model time for the whole task in Chrome, greedy. transformers.js runs Qwen3.5's linear-attention layers through a generic ONNX graph; its prompt processing is what makes each step take minutes.