LocalLLM Rigs

About Daniel Okafor-Lindqvist

user-content-Daniel Okafor-Lindqvist

Daniel researches and compares small, quiet machines that can run open-source language models at home, from unified-memory mini PCs to compact single-GPU builds. He prefers boxes with lots of fast shared memory over raw GPU compute. His focus is what a reader actually does with a machine: running a private coding assistant, summarizing documents offline, or keeping a 70B-class model available on a shelf without a jet-engine fan. He reads data sheets, independent benchmarks and Amazon price histories side by side so that the tokens/sec figures he cites are tied to a specific model and quantization. The spec he finds most overlooked is memory bandwidth, and the one he finds most overrated is the 'TOPS' number printed on the box.

How I choose and rate

I build every recommendation from three kinds of sources: Amazon listings and their price history, manufacturer data sheets, and independent reviews and benchmark measurements published on the web.

  • Usable memory: 32 GB is enough for 7–14B models at 4-bit. Plan on 64 GB or more if you want 70B-class models. This decides which models you can run at all.
  • Memory bandwidth: token generation speed scales with bandwidth. Around 250 GB/s is solid for a mini PC, and 500 GB/s or more feels fast. This decides speed versus price.
  • Published tokens/sec: I only trust figures tied to a named model, quantization and context length. Roughly 10 tokens/sec is where chat starts to keep up with your reading speed. This separates real performance from marketing.
  • Software support: a box is only as good as its support for llama.cpp, Ollama, CUDA, ROCm or MLX. This decides between cutting-edge chips and setups that just work.
  • Power and noise: for a machine that stays on your desk all the time, you want an idle draw under about 15 W and a fan that stays quiet under load. This decides between compact and powerful.
  • Price per GB of usable memory: this is the fairest way to compare a mini PC with a desktop GPU build. It decides where your budget goes further.