Get your rig on the board.
One command. About two minutes.
Load a model
Any model, any quant, loaded in any local server.
Node 18 or newer
Run the bench
The CLI finds the server, reads which model and quant it is running, signs you in with GitHub the first time, runs a fixed test and posts your score.
npx toks-gg benchGet your board
Your rig lands on the board for that model and quant, with its rank and a share card.
Bench whatever you actually run.
Each model and quant gets its own board, so you're always compared like with like.
Ollama
Load the model you want to test (any model, any quant):
ollama run qwen3:32b-q4_K_M "hi"Bench it:
npx toks-gg bench
Bench options
Add --name "Your rig name" and --cost 1900 to show up in the $ per tok/s ranking.
| Option | What it does |
|---|---|
--model | which loaded model to bench, if the server has several |
--url | your server, if it isn't on a common port |
--backend | pick one when several are running (ollama, llama.cpp, vllm, mlx, ...) |
--tp | tensor-parallel size, for vLLM across several GPUs |
--name | a name for your rig, e.g. "The Space Heater" |
--cost | what the rig cost you, for the $ per tok/s ranking |
--dry-run | run it, don't submit |
What gets sent
Nothing else. No prompts or files from your machine.
| Item | Exactly what is sent |
|---|---|
| GitHub username | From the sign-in, shown on your rig page. |
| GPUs | Their names as nvidia-smi, rocm-smi or macOS report them, plus the PCIe link on NVIDIA. |
| CPU | The model name. |
| RAM | Total size in GB. |
| Operating system | Its name and version. |
| Server | The backend's name, and the tensor-parallel size if you pass --tp. |
| Model | The name the server gives it, plus its file name, quant and file size when the server reports them. |
| Speeds | Generation and prompt processing, in tok/s. |
| Output hash | A SHA-256 hash of the model's output, not the output itself. |
| Run details | The one-time challenge ID and the CLI version. |
| Rig name and cost | Only when you pass --name or --cost. |