Updating the runtime

What the update does

llama.cpp ships ~weekly. Each release usually includes:

• Bug fixes (the most important; sometimes a release fixes a serious generation bug)
• New model support (architectures, attention variants)
• Performance improvements (faster kernels for specific GPU generations)
• Feature additions (new flags, new sampling methods)

The Runtime card in the Maintain tab compares your installed version against the latest available on GitHub and shows you the gap.

How to update

1. Maintain → Runtime card.
2. If a newer version is available, the card shows current vs latest and an Update runtime button.
3. Click. Blueprint:
   • Downloads the new build for your OS
   • Verifies the binary launches
   • Replaces the old binary atomically (writes to .new then renames, so a failed download doesn't leave you with no runtime)
   • Updates the version in the card

The Service / supervised llama-server keeps running on the OLD binary until next restart. Click Restart managed server in the Service card to swap to the new one.

Why you'd hold back on updating

• Pinned to a specific version for reproducibility: if your eval numbers are from llama.cpp build b4321, switching to b4500 may shift them. For production with a frozen eval baseline, pin intentionally and update only when you've re-validated.
• A recent release introduced a regression: occasionally a new llama.cpp build is slower than the previous one on certain hardware. Watch the llama.cpp release notes and the issue tracker for known problems before updating.

Common pitfalls

• Update during active serving: the update itself doesn't interrupt the running server (atomic file replace), but if you then restart, expect a model reload delay (multi-GB file from disk again).
• Disk space: the download is small (~150-300 MB), but combined with the old binary briefly on disk during the rename, plan for ~500 MB temporary headroom.
• Reverting: there's no built-in downgrade button. If a new release breaks something, the workaround is to grab the previous release manually from llama.cpp's releases page and drop the binary into ~/.blueprint/runtime/.

Updating the desktop app itself

Distinct from updating the runtime — the desktop app updates by downloading a new installer from llmblueprint.ai/download. See the release notes for what changed between versions.