What is a 70B model?
"70B" means a model has about 70 billion parameters (weights). It is a common size tier for open models and is noticeably more capable than 7B–30B models, but it is demanding to run at home. At full 16-bit precision the weights alone take about 140 GB, which is beyond every machine in this catalog. Quantized to 4 bits, they take roughly 40 GB, which is why reviews here treat 64 GB as a practical minimum and 96–128 GB as comfortable. The MINISFORUM UM890 Pro, with up to 96 GB of RAM, fits a quantized 70B model. The Apple Mac mini (M6), capped at 32 GB, does not. Fitting is only half the story: every generated token reads the whole model, so memory bandwidth largely decides how fast a 70B model responds.