I’m considering running a DeepSeek model locally and comparing a few quantization levels to fit the hardware I have. For everyday coding and reasoning prompts, where do people first notice a meaningful drop in quality?

If you’ve compared versions on the same machine, what tasks made the difference easiest to spot? I’m trying to choose a practical balance rather than judge the models only by a benchmark score.