AI, running on your hardware.

Step 1. Splitting your text into tokens

A model can't read letters or words. It reads tokens, which are fragments of text from a fixed list the model was built with. Common words are a single token, and rarer words get split into several pieces. Every token has an identity number, and those numbers are what actually goes into the model.

Type in the box and watch your sentence get split up. Each chip is one token, and the small number on it is that token's real id.

This runs against a real slice of the Qwen3 vocabulary, 30,747 of its 248,320 tokens, so the ids you see are genuine. For example " world" is one token, id 1814. Two limits worth stating: it's only a slice, so a word missing from it gets split into smaller pieces than the real model would use, and it matches the longest fragment first rather than following the model's exact merge order.

Step 2. Loading the model into memory

Before it can answer anything, the whole file of weights has to be copied from your drive into memory, and it has to stay there for as long as you're using it. This is the step that decides whether you can run a model at all. If it doesn't fit, nothing else matters.

The bar shows how much memory the current model needs against how much the selected hardware has. The first segment is the graphics card, and if the model is too large to fit there, the overflow spills into ordinary system memory.

…

Step 3. Reading your question, then writing the answer

Running a model happens at two very different speeds, and this is the most useful thing on the page to understand. First it reads everything you sent. Because that text already exists, the model can work through all of it at once, which is fast. That stage is called prefill, and it's why there's a pause before the first word shows up.

Then it writes. It can only produce one token at a time, because each one depends on the one before it, so the answer arrives as a steady trickle rather than all at once. That stage is called decode, and its speed is what you actually watch on screen.

Compare the two numbers below. The gap between them is the point: reading is cheap, and writing is what costs you time.

…

…

Step 4. Remembering the conversation

As the model reads, it works out two things about every token it has seen: a key, which is how that token gets found again later, and a value, which is what that token contributes once it's found. Redoing that for the whole conversation every time it writes a new word would be enormously wasteful, so it keeps them in memory instead. That store is the KV cache, named for keys and values.

It matters because it grows with every token in the conversation, and it sits in the same memory as the model. A long conversation can end up needing more space than the model itself.

…

…

…

…

Step 5. Choosing the next word

The model never picks a word. What it produces is a probability for every token in its vocabulary, tens of thousands of numbers saying how likely each one is to come next. Turning that list into a single choice is called sampling.

Always taking the most likely token makes text that's flat and repetitive, so a controlled amount of randomness gets added, tuned by settings called temperature and top-p. That's why asking the same question twice can give you two different answers.

Below is what the model might produce after the words "The cat sat on the". These ten candidates are an illustration, not a real model's output, because this site models memory and speed rather than running a model in your browser. The two controls, however, do exactly what they do in real software, so the way the bars reshape is accurate.

…