My hunch is that this might not just be reasoning distillation; it could even be benchmark distillation. This is a small experiment and definitely doesn’t prove anything about how any model was trained, but I thought the contrast was interesting enough to share.
I’d also like to try glm-5.3, but it isn’t fully available across the model-serving platforms I use yet. If I get around to running it and the results are interesting, I’ll share them later.