Months of hands-on testing with locally run large language models (LLMs) show that raw parameter count is less important than architecture, context window, and memory bandwidth. Advances in ...