Summary
This video takes a close look at Dory, a 4 billion parameter looped transformer, also called a latent recurrent transformer, released by NVIDIA’s Nemotron Labs. Because the release comes with little more than a model card, the presenter installed the model and built a custom UI to show how it behaves in practice.
The core idea is that a single set of weights can run a middle recurrent block multiple times per token. Dory splits its layers into a 16-layer input block, a 40-layer recurrent block, and a 16-layer output block. At one loop the model has about 72 layers of effective depth, at two loops 112, and at four loops 192, giving roughly the compute of a 9B dense model with the weight memory of a 4B model. NVIDIA’s composite score puts it ahead of Gemma 4 26B.
The video also covers the tradeoffs. Each loop keeps its own KV cache by default, so memory grows with loop count, though a last-loop mode keeps a single cache and cuts memory for long-context, high-concurrency serving. Side-by-side runs at one, two, and four loops show differences in speed, token count, and answer quality, though more loops are not a guaranteed improvement. Model specs include 256K context, pretraining on 6 trillion tokens with the first 4.5 trillion in NVFP4, and release under NVIDIA’s open model license with multiple quantized variants.
📺 Source: Sam Witteveen · Published November 02, 2026
🏷️ Format: Deep Dive
