Run

Run a benchmark

One container. One command. No account. No email address.

Read the code first

LocalMax runs on your machine. Examine the source before you run it. Every part is public: the runner, the container files, the profiles and the checks.

Collected

GPU model, VRAM, driver, CUDA, PCIe link, power limit and clocks. CPU model and core count. RAM. OS and kernel. All measured values, the raw records for each request, and the telemetry.

Never collected

Hostname. Username. File paths. GPU serial numbers. Board UUIDs. MAC addresses. Environment variables. Data about other processes. The runner removes secrets and paths from logs on your machine, before it writes them.

Your name is generated

LocalMax gives your machine a name such as Rocinante-K7X2P. The name comes from a key on your machine. You do not type a name. Keep the 5-character code. Use it to find your results at /systems/CODE.

Submission is voluntary. You control what you send. Commandlocalmax inspect shows the full content before upload. Legal · Privacy

1. Check your system

You need Linux, the NVIDIA driver and the NVIDIA Container Toolkit. LocalMax supports x86_64 and aarch64. WSL2 works, but it is experimental. LocalMax marks a WSL2 result.

# 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

Select the highest tier that your total VRAM permits. Then run a lower tier also. This makes your system comparable across the full range.

Profile	Tier	Min VRAM	Model	Download	Run time
diffusion-enthusiast-fp8	Enthusiast · FP8	24 GB	8B fp8_e4m3	9.3 GB	~14 min
diffusion-entry-fp8	Entry · FP8	12 GB	4B fp8_e4m3	3.8 GB	~14 min
diffusion-prospector-fp8	Prospector · FP8	64 GB	12B fp8_e4m3	13 GB	~14 min
llm-enthusiast-fp8	Enthusiast · FP8	24 GB	8B fp8_e4m3	8.8 GB	~14 min
llm-entry-fp8	Entry · FP8	12 GB	4B fp8_e4m3	4.1 GB	~14 min
llm-prospector-fp8	Prospector · FP8	64 GB	32B fp8_e4m3	35 GB	~14 min
vision-enthusiast-fp8	Enthusiast · FP8	24 GB	8B fp8_e4m3	9.3 GB	~14 min
vision-entry-fp8	Entry · FP8	12 GB	4B fp8_e4m3	4.9 GB	~14 min
vision-prospector-fp8	Prospector · FP8	64 GB	32B fp8_e4m3	36 GB	~14 min

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 runtime and waits for it. Then it runs a warm-up. Then it runs the fixed workload and records telemetry. Stop all other GPU work first. The runner measures everything on the GPU, including your browser.

4. Inspect before you publish

Nothing goes to LocalMax until you accept it. Command inspect needs no network. It shows the manifest and each file for upload.

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

Collected

GPU model, VRAM, driver, CUDA, PCIe link, power limit and clocks. CPU model and core count. RAM. OS and kernel. All measured values, the raw records and the telemetry.

Never collected

Hostname. Username. File paths. GPU serial numbers. Board UUIDs. MAC addresses. Environment variables. Data about other processes. The runner removes secrets from logs on your machine.

5. Submit

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

The CLI shows a verification link. Open the link. Complete the check. The upload then starts. The check makes mass fabrication expensive. It does not identify you.

Your result appears in one or two minutes. LocalMax computes each metric again from your raw records before it accepts the result. Your result page lists each check.

If something fails

An out-of-memory failure is a valid result. Submit it. It records that the profile does not fit that hardware. The runner marks the workload as failed. It does not retry with different settings.

Report problems atgithub.com/pavel4ai/localmax.net/issueswith the run ID.