Hacker News Reader

Only if you double layers by layers instead of the whole stack (which IIRC is what nanbeige is doing).

To put it simply, if you have 3 layers A-B-C then A-A-B-B-C-C requires more compute but not more memory bandwidth, but A-B-C-A-B-C requires both twice the compute and twice the memory bandwidth for the same token generation speed.

It's the opposite kind of bundling as MoE, which effectively trades memory/storage to save on compute at inference time.

It’s approximately the same as Qwen3.827b’s propensity to think a lot, right?

Not quite. Looped models do the extra "thinking" inside the model's layers. So the token gets twice the number crunching performed on it before it gets spit out. I think of it as the first loop "kickstarts" the process, and the second loop refines it.