Llama Nemotron Rerank
NVIDIANVIDIA's NeMo Retriever reranking model, a 1B cross-encoder fine-tuned from Llama-3.2-1B with bidirectional attention and a binary relevance head. It supports 8,192-token documents and was evaluated across 26 languages on MIRACL, MLQA and MLDR, and it is the reranking stage NVIDIA ships as a NIM microservice alongside its llama-nemotron embedding model. A vision-language sibling, llama-nemotron-rerank-vl-1b-v2, extends it to visual documents.
Added 2026-09-11 for the embeddings_retrieval promotion. The Hub licence field reads `other`; the card resolves it to the OpenMDW License Agreement v1.1 with the Llama 3.2 Community License named as additional information. Neither name appears in the pretrained openness ladder, so the formula abstains here - see the score note. Verified 2026-09-11 via the llama-nemotron-rerank-1b-v2 model card.
Openness
3 medium confidence- weights
- open(ungated safetensors for the text reranker
- data
- closed(800k samples from an unnamed blend of commercially licensed public QA datasets
- code
- partial(card usage and NIM deployment only
- license
- OpenMDW-1.1(the governing licence named in the card)+Llama-3.2-Community(named as additional information
The Hub licence field reads `other`; the card resolves it to the OpenMDW License Agreement v1.1 with the Llama 3.2 Community License as additional information, and the shared ladder now names both. OpenMDW-1.1 imposes no cap on who may use the weights or at what scale and sits at permissive_non_osi; the Llama 3.2 Community License carries the same 700M monthly-active-user bound as the 3.1 licence beside it and sits at use_bounded, which is the more restrictive of the two and the one multi_sku_rule lets decide. Confidence is medium because the governing terms are spread across a bare `other` tag and a prose section of the card rather than declared once.
- https://huggingface.co/nvidia/llama-nemotron-rerank-1b-v2/raw/main/README.md recorded 2026-09-11
Front matter reads license: other, license_name: openmdw-1.1, license_link https://openmdw.ai/license/1-1/. The License/Terms of use section says use is governed by the OpenMDW License Agreement version 1.1, with Additional Information: Llama 3.2 Community Model License Agreement; Built with Llama. The card states the model is ready for commercial use, describes a fine-tune of meta-llama/Llama-3.2-1B, and records training on 800k samples from an unnamed blend of public QA datasets chosen for commercial licensing.
Adoption
4 high confidence1,052,707 downloads in the trailing 30 days for llama-nemotron-rerank-1b-v2, the text reranker. The vision-language sibling is excluded under the same rule qwen3-reranker and nemotron-embed follow - a VL line is a separate product from the text line - and it would add about 26K without moving the band. Just over the 1M threshold on the text model alone.
- https://huggingface.co/api/models/nvidia/llama-nemotron-rerank-1b-v2 recorded 2026-09-11
downloads: 1097075; gated: false; private: false for nvidia/llama-nemotron-rerank-1b-v2.
Capability
3 medium confidenceRung 3, solid workhorse: credible published numbers, real multilingual and long-document coverage, and a production NIM path, but the evaluation is a pipeline Recall@5 rather than a standalone retrieval-quality figure comparable to the frontier, and NVIDIA own table puts the older nv-rerankQA-mistral-4b-v3 pipeline ahead of it at 75.45%. One rung below qwen3-reranker, which publishes frontier BEIR and MMTEB numbers on the instruments this category treats as primary. Confidence is medium because no public leaderboard carries this model.
- https://huggingface.co/nvidia/llama-nemotron-rerank-1b-v2/raw/main/README.md recorded 2026-09-11
Evaluation tables report llama-nemotron-embed-1b-v2 + llama-nemotron-rerank-1b-v2 at 73.64% average Recall@5 on NQ, HotpotQA, FiQA and TechQA against 75.45% for nv-embedqa-e5-v5 + nv-rerankQA-mistral-4b-v3, and 65.80% on MIRACL multilingual; the overview states support for 8192-token documents and evaluation on 26 languages.
Verified 2026-09-11