I’m comparing local quantized DeepSeek models for coding help, and I’m trying to decide whether a smaller build is a reasonable trade-off or a noticeable drop in answer quality. General benchmark scores don’t always tell me how the model handles debugging or multi-file changes.

For people who have tested several quantization levels, what kinds of coding tasks showed the biggest difference? I’d especially value comparisons using the same prompts and evaluation set, rather than impressions from a handful of chats.