Chat Visualiser.
Chat with a small instruction-tuned language model and watch its reply form one token at a time. Your message and the model’s answer are split into tokens and placed in a 3D space by their embeddings: your tokens in amber, the reply in teal.
After each reply, click any token to see which earlier tokens the model attended to, or switch the transcript to probabilities to see how likely each reply token was and which alternatives the sampler passed over. Step through the conversation token by token, or follow it through the model’s layers. The model has no inner voice to reveal; this is a look at the attention, learned patterns and sampling behind a chatbot’s answer.
Not signed in. Sign in to use your Hugging Face ZeroGPU quota. On the Space page, sign-in opens the app in its own tab, where you stay signed in.
What you will see: your conversation as a 3D map of tokens, with attention links and next-token probabilities for every reply. Small instruction-tuned models work best; custom models need a chat template.