GPU benchmarks for local LLMs

Real Ollama inference results across 23 GPUs and 16 models. Updated Mar 13, 2026.