Install quickstart

What you need

• A Mac with Apple Silicon and enough free disk space for a model artifact and runtime files.
• A current InferGrade Hub browser session.
• The signed and notarized Desktop Runner DMG from the public InferGrade Runner release.
• Hugging Face access only when you choose a gated or private artifact. Public models do not require it.

The Build Week submission remains compatible for judging. Use the latest Runner for newer model and runtime support.

1. Install Desktop Runner

1. Open the InferGrade Runner DMG.
2. Drag InferGrade Runner into Applications.
3. Launch the app and allow macOS to open it if Gatekeeper asks for confirmation.
4. Keep the app open while you pair and run the first benchmark.

2. Pair the Runner

1. Open Hub Setup.
2. Sign in, then create a one-time pairing code.
3. Paste the code into Desktop Runner.
4. Wait for Hub Setup to show the paired Runner as ready.

3. Install or select the runtime

1. In Desktop Runner, install or select the managed llama.cpp Metal runtime.
2. Use the Runner self-test before starting a benchmark.
3. If runtime setup fails, use the support export and recovery guidance before retrying.

4. Queue one known-good run

1. In Hub Recommend, choose a known-good question or open a public Result follow-up.
2. Use Build to prepare one local run from that recommendation.
3. Queue the run and keep Desktop Runner open so it can claim, execute, and upload.
4. Open the Result page to inspect what the run proves, what it does not prove, and the next benchmark.

If something blocks

• Pairing code expired or already used: create a fresh code in Hub Setup. Do not paste old codes into issues, chats, or screenshots.
• Runner token revoked or expired: unpair in Desktop Runner, then pair again with a new Hub code.
• Runtime setup failed: retry the managed runtime install, select an existing llama.cpp binary, or create a support export before changing paths.
• Model path missing or gated: choose a local GGUF file or reconnect Hugging Face access, then run the self-test again.
• Upload rejected after a local run: keep the local artifacts, do not edit run ids by hand, and use the support export for recovery.