TokenBench protocol
Benchmark your card
One single command. It downloads llama.cpp (official release) and the reference model (~4.9 GB, once, cached in ~/.tokenbench), runs the standard protocol, collects telemetry, then publishes the result.
Download our app
Model library, live telemetry, sustained tests with curves: download, double-click, benchmark.
Windows SmartScreen may warn on first launch (unsigned app): "More info" then "Run anyway".
Don't want to use the app? Follow these instructions
Prerequisites
Node.js LTS and an NVIDIA GPU with up-to-date drivers. That's it.
Windows: once, before the first launch
PowerShell blocks scripts by default (npx included). Allow them for your account — no admin rights needed:
Set-ExecutionPolicy -ExecutionPolicy RemoteSigned -Scope CurrentUser
The command
Windows (PowerShell), macOS or Linux:
npx https://tokenbench.fr/tokenbench-0.4.2.tgz --token 'VOTRE-JETON'
The token is OPTIONAL: without it, the command works the same (just drop --token) and your measurement lands "pending moderation". With an API key from your account (My account → API keys), the run is published under your username. Single quotes around the key: special characters don't survive the shell without them.
Benchmark another model
Beyond the protocol, any model in the catalog (Qwen, Llama, Gemma, Mistral, DeepSeek…) can be measured: find its identifier, then run it. The measurement is automatically linked to its page on the site. Browse the catalog →
# 1. Find the identifier (and check that the model fits on your card) npx https://tokenbench.fr/tokenbench-0.4.2.tgz --list-catalog qwen3 # 2. Download it, measure it and publish the result npx https://tokenbench.fr/tokenbench-0.4.2.tgz --token 'VOTRE-JETON' --catalog qwen3-8b
The file (Q4_K_M quantisation) is downloaded once into ~/.tokenbench; you are offered to delete it at the end. In the app, use the “Add a model” section of the library.
Useful options
--gpu-index 1 # multi-GPU machine: benchmark the 2nd card --ref-only # reference model only (fast) --list-catalog [mots] # list the catalog models (optional filter) --catalog <id> # benchmark a catalog model (e.g. gemma-3-12b) --replay result.json # re-publish a result whose upload failed --no-upload # measure without publishing --help # everything else
What gets measured
Generation and prompt processing speed (llama-bench, pp512/tg128, 5 repetitions), VRAM peak, temperatures and power draw (nvidia-smi, sampled during the run). Nothing is typed in by hand: everything in the database was measured.