Where might a model's reasoning branch? Each token is shaded by a heuristic: logit entropy (how unsure the model is about its next token), the outcome probe (a linear probe on layer-18 hidden states predicting how much forcing a different token would change the final answer), or their product. Click any token to see what else the model considered there.
This runs Qwen3-0.6B entirely in your browser: on your GPU via WebGPU where available (about 570 MB, downloaded once and cached), otherwise on your CPU (about 620 MB, and slower). Live mode shows logit entropy only: the outcome probe needs hidden states the browser build of the model doesn't expose, and it was trained on the larger model anyway. Decoding is greedy.