Isham Faizal published an interactive LLM Attention Visualizer on September 7, 2026. It shows a prompt and the text generated by a small language model. Hovering over a generated token brightens earlier tokens according to an aggregated attention signal, so readers can explore patterns in how the model uses previous text.
The browser runs Qwen3-0.6B locally through Transformers.js. Pre-generated examples appear immediately. To generate your own text, you download an instrumented model of roughly 618 MB. Faizal modified its ONNX computation graph to expose internal values normally unavailable to the browser runtime.
Faizal scales attention weights by the size of their value vectors, combines scores across heads and layers, then uses those values to set token opacity.
Stronger highlights can suggest where the model focused. They cannot establish why it generated a particular word: the visualization leaves out computations beyond the combined attention scores. The project is a way to explore this small model, not a complete explanation of its reasoning.