▦

Attention Heatmap

Type a sentence. A real Transformer (BERT) runs in your browser and shows its self-attention — which words each word looks at — across all 6 layers × 12 heads.

</> Source

Loading the Transformer (~91 MB, first time only)…

Layer

Head

View

Each row is a word doing the looking; brighter cells are the words it attends to. Hover a cell for the exact weight, or hover a token to isolate its row/column.

low

high attention weight (each row scaled to its own max)

Real self-attention read out from 🤗 Transformers.js (all-MiniLM-L6-v2, a BERT encoder). Model + inference run locally — nothing leaves your browser.