Anyone have benchmarks on how the llama 3 8b model performs when quantized to varying degrees? I reckon many people will be running these with llama.cpp or similar.