LLMs have been trending towards obscenely large number of parameters (314B for grok), which makes quantization crucial if you want to run them without a Meta-sized budget.