Skip to content

Token speed visualiser

Same answer, two speeds. Watch it arrive at each rate and see what a tokens-per-second number feels like in the room.

Set the race

Compare
tokens/s
tokens/s
1 to 2,048

Keys: P pause or resume, R replay.

The same answer, twice

Stream A

5tokens/s

ready
Emitted
0 / 60
Elapsed
0.0s
Finish in
12s

No tokens yet.

Stream B

50tokens/s

ready
Emitted
0 / 60
Elapsed
0.0s
Finish in
1.2s

No tokens yet.

Both streams show the same 60 tokens, taken in order from one fixed sample. Each chip is one real cl100k_base token. A chip with a thick left edge began with a space.

Output over time

Tokens emitted over timeStream A at 5 tokens per second reaches 60 tokens after 12s. Stream B at 50 tokens per second reaches 60 tokens after 1.2s.00.0s153.0s306.0s459.0s6012sTIMETOKENS
  • Stream A, 5 tokens/s, solid line and circle
  • Stream B, 50 tokens/s, solid line and square
  • Dotted lines show where each stream will finish

What to take from it

  • At 5 tokens/s, a 60-token answer takes 12s. At 50 tokens/s, it takes 1.2s.

  • What is a token?

    A small piece of text. In English it averages roughly three quarters of a word, though spaces, punctuation and common words vary. The sample starts “Marrow Tide” and is the opening of the 4K text in the context visualiser.

  • Try this

    Replay with 1 vs 5. The gap stops being a benchmark and becomes a pause you can feel.

  • This is a simulation

    Token number N arrives at N divided by the rate, and nothing else. Real models wait before the first token, and their speed moves during an answer. Nothing here calls a model.

  • Which tokenizer

    Tokens come from cl100k_base through js-tiktoken 1.0.21, split ahead of time. Another model would cut the same text differently.