Tokenspeed

A new interactive tool called Tokenspeed lets users visualize and compare LLM token-generation rates across four modes—code, text, think, and agent—by simulating speeds from 5 to 800 tok/s. It also demonstrates prefill latency, showing how long-context prompts delay first-token output, a metric often omitted from benchmark throughput numbers.

Every local-LLM benchmark reports throughput: "47 tok/s on an M3," "180 tok/s on a 4090," "500 tok/s on Groq." Unless you've actually watched tokens stream at those rates, the numbers are hard to internalize. This is the rendering. Four modes code — syntax-highlighted pseudo-code, the most common thing you watch stream out of an LLM. text — lorem ipsum prose, for the chat/answer case. think — dim-italic reasoning sentences alternating with code, mimicking a reasoning model thinking out loud. agent — alternating tool calls and code generation with processing pauses, simulating an AI coding agent. What to try Start at the default 30 and read along. Then hit 1 5 tok/s — Raspberry-Pi-class local model , 5 60 tok/s — typical hosted Claude or GPT , 7 200 tok/s — Groq territory , 9 800 tok/s — Cerebras-class, where the bottleneck is your eyeballs . Now switch between c and t at the same rate. The difference is striking — and intentional. Prompt processing Before a model emits a single output token, it has to read your whole prompt — the prefill pass. Open p , pick a context size, and the tool makes you sit through that wait before streaming starts, the same way it makes you feel tok/s. Prefill is much faster per token than generation, but a long context still stalls you: 64k tokens at 1,000 tok/s of prefill is over a minute of nothing. That delay is time-to-first-token, and it's the half of the latency story a throughput number never shows. What counts as a token This approximates BPE-style tokenization, not any vendor-specific encoder tiktoken , Claude's tokenizer, etc. — those disagree in the details anyway . Short words are often one token; longer identifiers split into chunks processUserInput → process + User + Input ; punctuation and operators usually count too. Code is more token-dense than prose, so the same tok/s can feel very different depending on what's streaming. The benchmark number is honest; the perceptual effect varies a lot by content type — which is the gap this tool exists to expose. English prose averages ~1.3 tokens per word, so 30 tok/s ≈ 23 words/s.