Guides

Everything we know about getting a local AI server working at home. Every claim links to where it came from, and any number we could not source is left out rather than guessed.

• Q4, Q5, Q8, FP16: what quantization actually costs you

  The numbers in every model filename. What bits-per-weight means, what you give up going lower, and which one to pick for the box you own.

  Updated 2026-08-16

• Buying used AI hardware safely

  The used market is where the value is, and where the traps are. What to check before you pay, what protection you actually have, and what to test the day it arrives.

  Updated 2026-08-15

• Setting up llama.cpp, Ollama, and vLLM

  Three runtimes, three different jobs. Which one to install, the commands that actually work, and the memory knobs that keep them from falling over.

  Updated 2026-08-15

• Multi-GPU: tensor parallel, layer split, and when NVLink matters

  Two GPUs do not automatically mean twice the speed. What each split mode actually does, which one to pick, and when your interconnect becomes the bottleneck.

  Updated 2026-08-15

• Power, cooling, and noise

  The constraints that bite after you have already bought the hardware: what the wall can deliver, what a passive card needs, and what it will sound like in the room you sit in.

  Updated 2026-08-15

• Rig builds by budget

  What to buy at four price points, chosen on memory capacity and bandwidth rather than on benchmark charts we cannot verify.

  Updated 2026-08-15

• What can you actually run at home?

  Memory capacity decides whether a model loads. Memory bandwidth decides how fast it answers. Everything else is a distant third.

  Updated 2026-08-15