I want to run one of the quantized R1 models locally, but I’m unsure where the useful balance is between model size, VRAM, and response speed. I’d mostly use it for coding questions and math, not as a production service.

If you’ve tried different DeepSeek R1 distill sizes on a single GPU, which setup felt capable without being painfully slow? Please include the GPU and quantization level if you can; those details make comparisons much more useful.