BENCH YOUR RIG
One command. About two minutes.
You need Node 18 or newer and a model loaded in any local server. 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 benchAdd --name "Your rig name" and --cost 1900 to show up in the $ per tok/s ranking. --dry-run runs it without posting.
Set up your backend
Bench whatever you actually run. Each model and quant gets its own board, so you're always compared like with like.
Ollama
- 1.Load the model you want to test (any model, any quant):
$ ollama run qwen3:32b-q4_K_M "hi" - 2.Bench it:
$ npx toks-gg bench
llama.cpp
- 1.Start llama-server with your GGUF:
$ llama-server -m your-model.gguf -c 8192 -ngl 999 - 2.Bench it. The CLI reads the file name and size to work out model and quant:
$ npx toks-gg bench
MLX (Apple Silicon)
- 1.Serve an MLX model:
$ mlx_lm.server --model mlx-community/Qwen3-32B-4bit --port 8000 - 2.Bench it:
$ npx toks-gg bench
vLLM, SGLang, LM Studio, TabbyAPI, KoboldCpp, Jan, llama-swap
- 1.Start your server as usual. The CLI checks the common ports and identifies it. On a custom port, pass --url:
$ npx toks-gg bench --url http://127.0.0.1:8000 - 2.Splitting one model across GPUs with tensor parallel? Pass the size:
$ npx toks-gg bench --tp 4
What gets sent
Your GitHub username, GPU, CPU and RAM size, operating system, backend name, the two speeds, and a hash of the model's output. Nothing else. No prompts or files from your machine.