An interactive AI teaching demo

Follow the next token.

A language model writes one token at a time. At each step, it assigns probabilities to possible next tokens. Try a short prompt, then walk through the model's actual choices and alternatives.

Qwen2.5 0.5B · open-weight model How this works

Start a trace

Give the model a beginning.

Prompt

Keep it short. Your words are sent to the hosted model when you select Generate.

Try:

Generate up to

Show alternatives

Sampling controls

At temperature 0 the model picks the most likely token. Higher temperatures make other choices possible. The bars always show the model's original probabilities.

Temperature

Top-p

Checking the model…

Explore the response

One decision at a time.

Your token trail will appear here. Start with a prompt to see what the model considered.

What do these numbers mean?

Each bar is a probability from the model's next-token distribution at that exact step. A token can be a word, part of a word, a space, or punctuation. These numbers are not the model's thoughts or confidence that a statement is true. Sampling can choose a token outside the alternatives shown here.

The model runs on a public Hugging Face Space. The app does not save prompts or responses, but your prompt is sent to that service to make a trace. See the privacy page for details.