How fast is 10 tokens per second?
A benchmark number is hard to feel. Play the same AI reply from 0.5 to 30 tokens per second and find the point where waiting stops getting in your way.
Feel 10 tokens per second
The same reply is replayed at the selected output speed. Change the slider or tap a preset to compare.
▌
What does tokens/s measure?
Tokens per second is the rate at which an LLM generates its answer after it starts responding. tok/s, tokens/s, and t/s are all used as shorthand; llama.cpp's official benchmark uses t/s. We write tokens/s because it is easiest to understand at a glance.
This demo uses about four English characters per token. Real token length varies by language, tokenizer, punctuation, and code, so it is a feel test rather than a benchmark.
Text streams faster than most people can comfortably follow.
Smooth for chat; waiting rarely interrupts your train of thought.
Usable, but longer answers make the generation time noticeable.
Fine for short answers; tedious for explanations or code.
You will spend more time watching the answer arrive than reading it.
Why the first word can still take time
Output speed is only one part of responsiveness. The app must first load the model and process your prompt; that delay is usually called time to first token. A model can generate quickly once it starts and still keep you waiting before the first word appears.
AICanRun result pages show estimated or measured output speed for a particular phone, model, and quant. Use this page to translate that number into a rough experience, then check your phone.