AI speed questions, answered

For reading, anything above about 10 tokens per second is already faster than you can follow — a person reads roughly 4 to 7 tokens per second. Past that, extra speed stops improving the reading experience.

Where high speed genuinely matters is automated work. An agent making chained calls, or a job processing thousands of documents, finishes proportionally sooner at 500 tokens per second than at 50. For a chat window, it makes no difference.

Adults read roughly 200 to 300 words per minute, which is about 4.2 words per second. This site uses 250 wpm as its reference point and expresses every result as a multiple of it.

That framing matters because tokens per second is meaningless to most people, whereas '3x faster than you read' is immediately intuitive.

Two different delays get confused. One is how long before anything appears at all; the other is how quickly words arrive once they start. A model can score well on the second and still feel sluggish because of the first.

Reasoning models make this sharper. They begin producing tokens quickly, but those tokens are hidden thinking you never see. The visible wait can be several seconds longer than the raw latency number suggests.

Yes. Modern browsers expose your GPU through WebGPU, which lets a compact model run entirely on your machine with no install, no account and no API key. This site does exactly that when you open it.

The trade-off is size. Browsers cap how much memory a tab can use, so models in the 1–3 billion parameter range are practical while larger ones are not. Weights download once and are cached afterwards.

No. The on-device test loads the model into your browser and generates locally. The prompt and the output never leave your machine, and it costs nothing to run because it uses your hardware rather than a server.

Cloud comparisons are different by nature: those send the fixed benchmark prompt to the provider you choose. That prompt is the same for every model and contains nothing personal.

Data-centre accelerators have far more memory bandwidth than a laptop GPU, and generation speed is bound by how fast model weights can be read from memory. A specialised provider can be thirty times faster at producing words.

Time to first word is a different story. A local model has no network round trip to pay, so it can begin responding about as quickly as a cloud service — sometimes quicker, depending on the route.

The gap between sending a request and the first piece of the response arriving. It reflects queueing, prompt processing and network latency rather than generation speed, which is why it is reported separately from words per second.

It is also the easiest metric to fake: a server that flushes an empty chunk on connect can report a latency that is really just network round-trip time. This site starts its clock only on the first chunk with real text.

On-device speed depends on GPU clock ramping, thermal state and background activity. Consecutive single runs on one laptop varied by about 25%, which is why the test discards a warm-up pass and reports the median of three.

Cloud results vary for a different reason — server queueing — and reasoning models vary again because the length of their hidden chain differs between runs.

No. Speed and quality are separate properties, and this site measures only speed. A fast model may give worse answers, and a slow one may be worth waiting for.

Speed matters most when it changes what you can do: interactive tools need low latency, agent pipelines need high throughput, and reading a chat reply needs surprisingly little of either.

The measurements are not sponsored, and no provider pays for placement in the results. The methodology is published in full, including the cases where a measurement is refused rather than reported.

It is also not a laboratory. Numbers are taken from your browser, over your network, at one moment in time — good for a fair comparison between models on the same run, not a certified benchmark.

The methodology page explains exactly what is measured and the three ways a speed test can mislead you. Or just run the test — it takes a few seconds and needs no account.

Canonical home: https://howfastai.com