Try a model, read the numbers

Pick a model and a card, set the prompt and the knobs, and see the answer with the memory it takes and the speed it runs at. Memory and speed are real estimates from the Cookbook's arithmetic; the answer itself is a sample, because this page does not run a model.

kracht-playground

Qwen3 8B

ModelRun onSystem promptTemperature 0.7Max tokens 256

Tool calling

sample answer

A GPU runs the same small calculation across thousands of cores at once, which is exactly the shape of the matrix multiplies a neural network is made of. That is why a model that would crawl on a CPU generates in real time on a GPU: the work is parallel, and so is the hardware.

Not a live model call

Speedest

~176 t/s

Memoryest

5.3 GB

First token

~180 ms

Tokens

Speed and memory are estimates from the model and card, and err upward. First-token time and token count are illustrative.

A real playground calls the model behind an endpoint and streams the answer back: the agent runtime, which is the platform under this. The sizing that tells you the model will run on your card, and how fast, is the part that works today.

Size the model