NVIDIA’s Nemotron-3-Embed-8B-BF16 has claimed the top position on RTEB, the multilingual Retrieval Text Embedding Benchmark, posting an average NDCG@10 score of 78.46 across 16 tasks. The result places it above competing open embedding models and establishes it as the highest-performing checkpoint in the Nemotron-3-Embed collection for dense retrieval quality.
NVIDIA’s Nemotron-3-Embed-8B-BF16 tops the RTEB benchmark with an NDCG@10 score of 78.46 across 16 tasks.
The collection consists of three open checkpoints: Nemotron-3-Embed-8B-BF16, Nemotron-3-Embed-1B-BF16, and Nemotron-3-Embed-1B-NVFP4. All three are transformer encoders trained with bidirectional attention masking and use mean pooling over token-level representations to produce final embeddings. Every checkpoint supports a maximum sequence length of 32,768 tokens, enabling long-context retrieval use cases.
Standardized query: and document: input prefixes are used across all models to separate query and document representations within retrieval pipelines.
The 8B model produces 4,096-dimensional dense embeddings, while the smaller 1B variants offer a reduced deployment footprint without fully sacrificing retrieval accuracy. Nemotron-3-Embed-1B-BF16 achieves an NDCG@10 score of approximately 72.38 on RTEB, which NVIDIA characterizes as a roughly 27% error reduction over a prior 1B predecessor. That margin positions the 1B model as competitive among open embeddings at comparable scale, with its model card reporting state-of-the-art results on text chunk retrieval benchmarks within its size class.
Performance extends beyond RTEB to additional benchmarks. On MMTEB Retrieval, the 8B model records an average NDCG@10 of 75.5, while the 1B model reaches 71.0, representing approximately a 28% error reduction over a prior baseline. On the LMEB benchmark, the 8B model achieves 64.4%, a figure NVIDIA describes as a new state-of-the-art for retrieval embeddings, with the 1B model reaching 61.5%.
Benchmark coverage spans multilingual text and code retrieval, confirming consistent performance across diverse domains and input types.
Multilingual and cross-lingual retrieval are explicit design priorities for the collection. Testing has been reported across 34 languages, including Arabic, Hindi, Japanese, Korean, Spanish, Swahili, Thai, Ukrainian, Urdu, and Vietnamese. The models are positioned for semantic similarity tasks alongside dense retrieval, broadening their applicability beyond standard document search pipelines.
All Nemotron-3-Embed checkpoints are evaluated as retrieval embedding models rather than generative systems, with benchmarks focusing strictly on dense retrieval quality. NVIDIA frames the collection as accuracy-first retrieval embeddings, outperforming many existing open embedding baselines at similar or smaller scales. The Nemotron-3-Embed-1B-BF16 was derived from Nemotron-3-Embed-3B through a process of pruning and distillation, reducing model size while preserving competitive retrieval performance.
The open release of all three checkpoints makes the models accessible for deployment across a range of retrieval infrastructure configurations, from resource-constrained environments using the NVFP4 quantized variant to higher-accuracy applications using the full BF16 precision models. The combination of benchmark leadership, multilingual coverage, and long-context support marks the Nemotron-3-Embed collection as a notable development in open retrieval embedding research.
