I’m comparing the smaller R1 distill models for coding help and trying to decide whether running one locally is worth the trade-off. Most comparisons I’ve found focus on general reasoning scores, which don’t tell me much about debugging or making a careful change across several files.

For anyone who has tested both a distilled model and full R1 on real coding tasks, what differences stood out? I’m curious about correctness, how often the smaller models miss constraints, and whether a particular distillation size felt like a sensible balance.