A real language model runs in your browser. It takes the start of a sentence, and in a few seconds of arithmetic chooses the next word. This show slows that down and draws every step from the model's own numbers: nothing here is staged.
Recorded runs of Qwen2.5-0.5B-Instruct (24 layers, 494 million knobs), trained by its makers to ask for tools. Every token, chance and attention number is exactly what the real model produced; the commands really run on a Linux inside this page (no internet, three files). They play straight away, on a phone too.
All four run the loop you are about to watch. They differ in size (layers, numbers per token, dictionary), in what they read, and in a few design choices (how word order gets in, whether the output table is the input table again). The chat model reads a hidden instruction and your message as a chat, then answers instead of continuing. Mamba, the design with no attention, is in the Explorer.
The sentence below is only a suggestion. Type your own: the whole show is computed fresh for it.
Tap anything on the stage for its numbers; drag, scroll or pinch the cloud, 3D shows the real three-dimensional map (drag to turn it); ? explains each screen and words maps the picture to the proper terms. Nothing you type leaves this page. The step-by-step version · the Explorer · code.
Tokens, spots in the cloud, listening back, layers of fixed knobs, a wall of lights, a weighted dice. One token at a time, from the same frozen numbers. Every frame you saw was drawn from them.
Every language model runs this same loop. The big ones just have more of everything: more layers, far more knobs, a bigger dictionary, and years of reading instead of children's stories. The step is the same; the scale is not.
Want the numbers themselves? The step-by-step walk-through shows every one; the Explorer shows the whole machine at once.