Sam Witteveen · 2026-11-02 · notable
Sam Witteveen — 'Dory - NVIDIA's Looped Transformer'
Sam Witteveen explains NVIDIA's Dory, an open 4B looped transformer that reruns a shared block of layers 1, 2 or 4 times per token, and tests it against Gemma 4 26B.
Sam Witteveen walks through NVIDIA's Dory, a 4B model that spends more compute per token by looping its middle layers.
What is it?
'Dory - NVIDIA's Looped Transformer' went up on Sam Witteveen's channel on 2 November 2026. The video covers Nemotron Labs Dory 1.0 4B, a research release NVIDIA put on Hugging Face on 16 October 2026, with a live demo of the model at 1, 2 and 4 loops.
How does it work?
Dory splits its layers into an 8-layer input block, a 20-layer shared recurrent block and an 8-layer output block, and runs the shared block several times per token. The video explains this loop design, compares a per-loop KV cache with a last-loop cache that keeps memory flat, and shows how to deploy the model.
Why does it matter?
A looped model lets one checkpoint trade speed for accuracy at serving time, with no second model to download. The video shows where that pays off and where Dory still loses to larger models.
Who is it for?
developers running small local models