SYLIN

LLM Attention Visualizer

by ishamf · Tool · Text & code

Show HN: interactive page runs Qwen3-0.6B in the browser and highlights which past tokens influenced each generated token

AccessHosted appRuns onWeb

Evidence

  1. Sep 8
    Opacity per past token = attention weight times value-vector magnitude, aggregated over heads and summed over layers; author calls it a heavy simplification
  2. Sep 8
    Commenter finds it simpler than BertViz for understanding inference and a good starting point for learners

Alternatives

Works with