AI Mini-PC Database

I track these boxes because a 1-to-2 liter machine that idles near silent is, for a lot of people, the most practical way to run a model at home. The trade-offs are real, and they are not the ones marketing pages lead with. The headline NPU TOPS number tells you how fast the on-chip accelerator can run small vision and audio models. It does not tell you whether a 7B or 13B language model will fit, or how fast it will decode. For that the two numbers that actually decide things are the RAM ceiling and the memory type.

Two camps split this table. Apple's unified-memory boxes (the Mac mini and Mac Studio rows) share one LPDDR pool between CPU and GPU, so a higher-RAM config directly buys you a larger model you can load. The Strix Halo and NPU-equipped x86 boxes pair a discrete NPU rating with conventional or soldered RAM, and there a high TOPS figure can sit next to a memory ceiling that caps what you can realistically serve. I list both the stock and maximum RAM, the memory type, form factor, and TDP on every row so you can read past the one number a spec sheet wants you to fixate on.

Where I have validated it, each row carries a short list of models the machine can actually run, from Llama 3 8B and Mistral 7B down to Phi-3 Mini and Whisper. Click any name for the full spec sheet. The Amazon links below are tagged affiliate links; using them costs you nothing and helps fund the testing.