A personal comparison gave Gemma 4 E4B, Ministral 3 in 3B and 8B sizes, and Phi-4 Mini six questions about a CSV mixing currencies, units, and missing values. All used llama.cpp, Q4_K_M quantization, and a 16K context with tools, search, and reasoning disabled, but trial counts and sampling settings differed by model. None identified the cheapest 16 GB GPU, and some confused currencies or invented historical prices. Results from one table and quantization setup cannot establish general performance. Row selection, summation, and questions that cannot be answered should be validated separately.