Overview
nvidia open embedding collection for production rag, agentic-retrieval, code search, and agent memory. Released July 16, 2026 (Hugging Face blog + model cards), OpenMDW-1.1, 32,768-token context, ~34 languages.
Checkpoints
| Checkpoint | Role | RTEB (NDCG@10) | Notes |
|---|---|---|---|
| Nemotron-3-Embed-8B-BF16 | Flagship | 78.46 (#1 as of ~Jul 16) | Emb dim 4096; Ministral-3-8B base |
| Nemotron-3-Embed-1B-BF16 | Efficiency | 72.38 | Distilled via ModelOpt NAS |
| Nemotron-3-Embed-1B-NVFP4 | Blackwell | 72.00 | ~99.5% of 1B BF16; up to 2× throughput |
Recent Developments
- 2026-07-16: Tops rteb; also strong on mmteb Retrieval (~75.45 for 8B) and ViDoRe-V3 text (nvidia-nemotron-3-embed-rteb)
- NIM microservice for 1B; NeMo AutoModel fine-tune/distill recipes; vLLM
/v2/embedsupport