Hard to fake. Easy to spot.
Any number produced on your own machine can be tampered with.
toks.gg doesn't pretend otherwise. It makes faking hard, and makes fakes stand out.
The benchmark is fixed.
The toks-gg CLI runs a set test, about 7k tokens in and 256 out, on whatever model your server has loaded. It reads the model and quant from the server itself, and the file size when it can see the file. Each model and quant has its own board. You don't type numbers in. Prompt processing and generation are measured separately, because they're very different speeds.
Every run gets a fresh challenge.
Before each run, the server issues a one-time challenge: a random seed and prompt that only works once, for your account, within the hour. An old result, or someone else's, can't be replayed.
Hardware is detected, not typed.
The CLI reads your GPUs, CPU and memory, and the server looks up each GPU's memory bandwidth itself. Rig cost is the only thing you enter by hand, and it's labelled as owner-reported.
The physics check.
Generating one token means reading the active weights from memory once. So generation speed can't beat memory bandwidth divided by the bytes read per token. For mixture-of-experts models that's only the active experts, worked out from the model's parameter counts and the real file size.
A 4060 claiming 140 tok/s on a 70B model is flagged automatically, because the hardware can't do it.
What the badges mean.
Every run on the board carries one of these.
| Status | What it means |
|---|---|
| Self-reported | Ran through the CLI, but we don't know that model's parameter counts or that GPU's bandwidth yet, so it can't be physics-checked. |
| Verified run | Came from the CLI and passed the physics check. |
| Reproduced | Another rig with the same hardware landed within 10%. |
| Over the ceiling | Faster than the hardware can go. Flagged automatically and struck through on the board. |