NVIDIA's New Embedding Models Deliver Massive Score Boosts
tomaarsen · x · 2026-07-17
NVIDIA's Nemotron-3-Embed-1B utilizes an unusual compression pipeline: starting from a 3B parent, it undergoes NAS pruning to 2B via NVIDIA ModelOpt, followed by distillation to 1.14B using an 8B teacher with cosine + MSE loss.
The author shares retrieval evaluation results: on 16 RTEB tasks, Nemotron-3-Embed-8B achieved an average NDCG@10 of 78.46, while the 1B version scored 72.38. The previous generation llama-nemotron-embed-1b-v2 scored only 60.47 at the same size, indicating an 11.9-point improvement for the 1B model. Similar gains were observed on MMTEB Retrieval (71.04 vs 59.58).
Related event: NVIDIA launches Nemotron-3-Embed retrieval models(11 posts)→
More from Models
- ICML 2026 position paper: LLM 'thinking traces' are a harmful anthropomorphism — rao2z · 2026-10-27
- One stray 'u' and Claude starts spelling like a Brit — cantrell · 2026-10-27
- Cohere's Invent-A-Dataset paper: zero-seed synthetic data beats frontier APIs on quality and diversity — idanbeck · 2026-10-27
- ARC Prize redesigns leaderboard: highest reasoning effort by default, CSV export — mhmazur · 2026-10-27
- Cohere: open-weight Qwen3.8-Flash-Next tried uploading probe file to Beijing cloud bucket in sandbox — wstone_bd · 2026-10-27
- Model refuses wrong psychiatric advice in deliberately misleading test — FellMentKE · 2026-10-27