NVIDIA's Nemotron Labs Dory: optimized 1B model rivals 30B models on reasoning
yaroslavvb · x · 2026-10-20
Sharing NVIDIA Nemotron Labs' new Dory collection of local-AI-optimized checkpoints on Hugging Face, yaroslavvb notes he long suspected advanced optimization methods would benefit small models most, since that's a harder optimization problem. The chart backs it up: Dory-1B scales from 13.8 to 18.4 on an AA v4.32 estimate as inference width goes k=1→4, beating Qwen3.5-9B (12.53) and Gemma4-26B-A4B (16.35), and approaching Muse-Glimmer-30B (19.94) and Qwen3.6-35B-A3B (20.15) — a 1B model rivaling 30B-class peers.
More from Infra
- 26-author study: a hardwired frontier AI training pause is technically feasible — wfithian · 2026-10-20
- Stardock CEO: local AI will dominate most AI work in 18 months, datacenters pointless — draginol · 2026-10-20
- Students get 1 year of Google AI Pro free, with 35 hours of monthly A100 on Colab — CeFurkan · 2026-10-20
- 768p video with audio generated on just two desktop DGX Sparks using new FastVideo QAD-v2 — Scobleizer · 2026-10-20
- AI power shortage is overstated: upgrades and exports shrink the gap, glut by 2028 — toptickcrypto · 2026-10-20
- Nebius launches live GPU spot pricing tracker for H100/H200/B200/B300 — demian_ai · 2026-10-20