ByteDance-linked looped-depth models win under matched FLOPs, params and KV cache

georgejrjrjr · x · 2026-09-03

The author corrects the record on looped-depth model research: ByteDance's Parcae paper already showed looping is compute-optimal under isoFLOP and isoParameter settings, and the new work is the first to establish the finding with FLOPs, parameters and KV cache all matched.

On "why not just make the model deeper?", the author's guess is that lower token capacity aids generalization in the data-limited regime, and the finding that compute efficiency barely suffers makes looped architectures a bullish direction for local models.

Related event: ByteDance Research Suggests Looped Transformers Beat Deeper Models in Compute Efficiency(2 posts)→

Original post →

More from Models

Models channel →