Run a benchmark

1. Check your system

Linux with the NVIDIA driver and Container Toolkit installed. x86_64 or aarch64 — DGX Spark is supported natively. WSL2 works but is marked experimental, and results from it are labelled as such.

# Reports driver, CUDA, GPU, disk, RAM and which profiles you qualify for
docker run --rm --gpus all ghcr.io/pavel4ai/localmax-llm:latest doctor

2. Pick a profile

Pick the highest tier your total VRAM supports. Running a lower tier as well is useful — it is how a system gets comparable across the whole range.

INT4 and NVFP4 lanes exist for each tier and are listed on theprofile matrix. They fit a larger model in the same VRAM and are ranked separately.

3. Run it

# Weights are cached in ~/.localmax and reused across profiles
docker run --rm --gpus all -v ~/.localmax:/cache \
  ghcr.io/pavel4ai/localmax-llm:latest run llm-entry-base

# Vision and diffusion use their own images
docker run --rm --gpus all -v ~/.localmax:/cache \
  ghcr.io/pavel4ai/localmax-vision:latest run vision-entry-base

docker run --rm --gpus all -v ~/.localmax:/cache \
  ghcr.io/pavel4ai/localmax-diffusion:latest run diffusion-entry-base

The container starts the pinned inference runtime, waits for it to become ready, runs a warm-up, then drives the endpoint with the fixed workload while sampling GPU telemetry. Close other GPU work first — the runner records what it sees, including your browser.

4. Inspect before you publish

Nothing is transmitted until you accept it. inspect works entirely offline and prints the manifest and every artifact that would be uploaded.

docker run --rm -v ~/.localmax:/cache \
  ghcr.io/pavel4ai/localmax-llm:latest inspect LAST

5. Submit

docker run --rm -v ~/.localmax:/cache \
  ghcr.io/pavel4ai/localmax-llm:latest submit LAST

The CLI prints a short verification link. Open it, complete a one-click browser check, and the upload proceeds. That check is the only thing standing between an anonymous contributor and the leaderboard — it exists to make bulk fabrication expensive, not to identify you.

Your result appears within a minute or two. It is validated asynchronously: every derived metric is recomputed from your raw records before it is accepted, and the finding list on your result page tells you exactly what passed.