Tools
LLM speed simulator
Paste a text, set the tokens per second rate, the delay before the first token and a bit of jitter: the text appears as if it were streaming out of a model. Useful to calibrate an animation, stress a chat interface or show the gap between two speeds. Everything is simulated in the browser, no model is called.
How many tokens the model emits per second once it has started. Around 5 you read faster than it writes; past 60 the text outruns your eyes. The slider stops at 600, the field accepts more.
The dead time between sending the request and the first token landing. It covers queueing, prompt processing and network. This is what makes an interface feel slow even when the stream itself is fast.
How irregular the gaps between tokens are. At 0 the stream is a metronome, which no real model ever is. At 35 each gap varies by up to a third either way.
Target rate
25 tok/s
Measured
-
Elapsed
-
Full run
≈ 57.8 s