Offline distillation lifts throughput 41%

Offline knowledge distillation reached up to 41% higher throughput on one H200 while running about 29% faster per iteration than online distillation at near-identical training loss. The practitioners’ preprint caches a teacher’s top-K logits and fuses a chunked KL loss, which let the team train at four times the prior context—32,768 tokens—on a single GPU; its implementation is public, though a patent application is pending. For teams building smaller deployment models, this is a reproducible systems path to test before paying for more training memory, while the archive’s analysis of open-weight control premiums remains the caution: matching training loss does not establish matching downstream quality.