Nvidia ships Dory1-4B, a looped LLM that reuses 20 layers to rival much larger models
scaling01 · x · 2026-11-03
Nvidia recently released Dory1-4B, a "looped" LLM with 36 layers: 8 input layers, a recurrent block of 20 layers (looped 1, 2 or 4 times), and 8 output layers.
Per the benchmark charts:
- More loops mean higher scores: 13.8 (k=1) → 17.7 (k=2) → 18.4 (k=4) on AA v4.3.2, beating Qwen3.5-9B (12.53) and approaching Muse-Glimmer-30B (19.94)
- Output tokens drop sharply with looping: down to 0.50x on Terminal-Bench 2.1 and 0.40x on GDPval-AA-V2 at k=4
Weight reuse trades parameters for compute depth — a small model punching above its size.
Related event: NVIDIA releases Dory1-4B, a looped LLM reusing 36 layers to rival 9B models(2 posts)→
More from Models
- Particle's model lands on data analytics Pareto frontier, beating Haiku 5.5 — ycombinator · 2026-11-03
- ML4 Goes Open-Weight: Frontier Code Agent and Cyber Capabilities, Far More Compact Than Preview — _AndrewZhao · 2026-11-03
- OpenRouter: heavy users now median 6 model families, 2x in a year — sarahookr · 2026-11-03
- Killed by Google eulogizes AI Overview — 'not yet,' replies a user — gaganghotra_ · 2026-11-03
- DeepSeek's 'super agent' sparks debate: is the meat proxy era over? — gdechichi · 2026-11-03
- Only 3% of humans beat Claude Opus 5.5 at predicting what people click — teortaxesTex · 2026-11-03