# toks-gg

Benchmark your local AI rig and post a verified tok/s score to [toks.gg](https://toks.gg).

```
npx toks-gg bench
```

It finds your running backend (Ollama, llama.cpp, vLLM, SGLang, MLX, LM Studio, TabbyAPI, KoboldCpp, Jan, llama-swap, or any OpenAI-compatible server), signs you in with GitHub, runs a fixed test (an ~7k-token prompt, then 256 generated tokens), and posts the result.

## Options

```
npx toks-gg bench --model qwen3-32b-q4_k_m     # pick a benchmark model
npx toks-gg bench --url http://127.0.0.1:8000  # any OpenAI-compatible server
npx toks-gg bench --tp 4                       # vLLM tensor parallel across 4 GPUs
npx toks-gg bench --name "The Space Heater" --cost 1900
npx toks-gg bench --dry-run                    # run it, don't submit
npx toks-gg models                             # list benchmark models
```

## How results are verified

- The server issues a one-time challenge (seed and prompt) for every run, so results can't be replayed.
- The CLI checks your backend is running the benchmark model. If it can't, the run is self-reported.
- The server knows each GPU's memory bandwidth, and flags any generation speed that's physically impossible for your hardware.

Details: https://toks.gg/verify
