Get working on real work with local AI. Don't go crazy over AI. Stay Mostlysane.
Loading models…
Good for creative writing, code review, and general advice.
Download a prebuilt binary — no compilers or build tools needed. The package includes a helper script that downloads your model and starts the server automatically.
Extract the zip, then run the helper script:
Prefer building from source? Switch to "Build from Source" above.
Context: tokens — auto-sized to fit your VRAM.
Copy and paste this anywhere — paths are absolute. Models go in
~/AI/models/ by default.
One-time calibration per model. The profile is quantization-invariant — same file works for Q8, Q5, Q4 quants.