Compare GLM, Qwen, and Gemma local LLMs - benchmarks, context sizing, and when home GPU inference beats cloud token costs.