I’m testing local DeepSeek models and trying to understand the trade-off between a smaller quantized build and a higher-precision one. On short prompts, the difference may be hard to notice, but does DeepSeek quantization make a more obvious difference when the prompt has several constraints or a long set of instructions?

If you’ve compared versions on the same hardware, I’d be interested in the kinds of failures you saw: missed requirements, weaker recall of earlier instructions, or something else. I’m not looking for a specific benchmark so much as practical experience with real prompts.