I’m comparing quantized versions of a DeepSeek model for local use and want to understand the accuracy trade-off, not just the difference in speed or memory use. Has anyone tested whether lower-bit versions make more mistakes on multi-step math problems, especially when the final answer looks plausible?

If you’ve run a comparison, what did you use as a test set, and did the results vary much between model sizes or quantization formats? I’m trying to decide what level of quantization is acceptable for a workload where wrong answers matter more than a few extra seconds.