--- title: NVIDIA Nemotron 3 Embed 8B takes #1 on RTEB multilingual... description: NVIDIA released Nemotron 3 Embed, a family of three open embedding models (8B and two 1B variants), with the 8B model claiming the top spot on the RTEB... canonical: https://daily.dev/posts/nvidia-nemotron-3-embed-8b-takes-1-on-rteb-multilingual-retrieval-leaderboard-b6jrl3kp7 twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: NVIDIA Nemotron 3 Embed 8B takes #1 on RTEB multilingual retrieval leaderboard | daily.dev og:description: NVIDIA released Nemotron 3 Embed, a family of three open embedding models (8B and two 1B variants), with the 8B model claiming the top spot on the RTEB... og:url: https://daily.dev/posts/nvidia-nemotron-3-embed-8b-takes-1-on-rteb-multilingual-retrieval-leaderboard-b6jrl3kp7 og:image: https://api.daily.dev/og/posts/B6jRl3KP7.png og:image:alt: NVIDIA Nemotron 3 Embed 8B takes #1 on RTEB multilingual retrieval leaderboard og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # NVIDIA Nemotron 3 Embed 8B takes #1 on RTEB multilingual retrieval leaderboard **[Collections](https://daily.dev/sources/collections)** · 2 min read · 1 upvotes · 0 comments ## Summary NVIDIA released Nemotron 3 Embed, a family of three open embedding models (8B and two 1B variants), with the 8B model claiming the top spot on the RTEB multilingual retrieval leaderboard at 78.5%. The models feature a 32k context window, multilingual and code retrieval support, and are built by converting causal decoder architectures into bidirectional encoders with structured pruning and knowledge distillation. NVFP4 quantization for Blackwell GPUs delivers roughly 2x throughput over BF16. Fine-tuning recipes are available via NeMo AutoModel. NVIDIA positions better retrieval as a way to reduce downstream token costs in agentic and RAG pipelines, with IBM, Zoom, Palantir, and ServiceNow already evaluating the models for production use. ## Content NVIDIA released Nemotron 3 Embed this week, a family of three open embedding models that currently sit at the top of the RTEB retrieval leaderboard. The flagship 8B model scores 78.5%, and the two 1B variants offer cheaper alternatives for latency-sensitive deployments. ## What's in the release Three models, all with 32K context windows and support for multilingual text and code: - **Nemotron-3-Embed-8B-BF16** — the flagship, #1 on RTEB at 78.5% - **Nemotron-3-Embed-1B-BF16** — 72.4% on RTEB, a 27% reduction in error rate over its predecessor - **Nemotron-3-Embed-1B-NVFP4** — FP4-quantized for NVIDIA Blackwell GPUs, up to 2x the throughput of the BF16 variant while retaining 99%+ of its accuracy All weights, datasets, and training recipes are open, including recipes for finetuning on your own data via NeMo AutoModel. ## How they were built The 8B model starts from a Ministral-3B-8B-Instruct backbone. The causal decoder architecture is converted into a bidirectional encoder, which handles full-sequence retrieval better than a standard left-to-right decoder. From there, the model is trained with a contrastive objective on web-sourced and synthetic retrieval pairs, then finetuned on more curated multilingual and real-world retrieval data. The 1B variants went through a more involved process. NVIDIA first applied the same bidirectional recipe to a 3B model, then used structured pruning and neural architecture search (NAS) to shrink it to 2B, then distilled against the 8B model to recover quality. That entire cycle was repeated a second time to reach a 1B model that holds up well on benchmarks. ## Why retrieval quality matters for agents RTEB is worth paying attention to here because it focuses on real-world tasks and does extensive decontamination to avoid leakage — so scores are harder to game than on some other leaderboards. The practical argument for better retrieval is straightforward: when a retrieval model returns irrelevant context, an agent has to search again. More searches mean longer reasoning chains and higher token costs. Tests with NVIDIA's Nemotron 3 Ultra showed that using these embedding models reduced the number of searches needed and cut downstream token costs measurably. Web search agents in particular showed clear efficiency gains. Enterprise partners including IBM, Zoom, Palantir, and ServiceNow are already evaluating the models for production RAG and agentic retrieval workloads. ## Questions this post answers ### What score does NVIDIA Nemotron 3 Embed 8B get on the RTEB retrieval leaderboard? Nemotron 3 Embed 8B (BF16) scores 78.5% on the RTEB multilingual retrieval leaderboard, taking the top spot. It supports a 32K context window and multilingual text and code retrieval, and is built on a Ministral-3B-8B-Instruct backbone converted into a bidirectional encoder trained with a contrastive objective. _Teams choosing an embedding model for RAG can track daily.dev for how retrieval benchmarks like this evolve._ ### What is the difference between Nemotron 3 Embed 1B BF16 and 1B NVFP4 variants? The 1B BF16 variant scores 72.4% on RTEB, a 27% error-rate reduction over its predecessor, while the 1B NVFP4 variant is FP4-quantized specifically for NVIDIA Blackwell GPUs, delivering up to 2x the throughput of the BF16 version while retaining over 99% of its accuracy. Both target latency-sensitive deployments. _Engineers weighing latency versus accuracy tradeoffs in embedding models can follow this comparison on daily.dev._ ### How were the smaller Nemotron 3 Embed 1B models trained from a larger model? NVIDIA first applied a bidirectional encoder conversion recipe to a 3B model, then used structured pruning and neural architecture search to shrink it to 2B, then distilled against the 8B model to recover quality. This entire pruning-and-distillation cycle was repeated a second time to produce the final 1B model. _Practitioners exploring model compression techniques for embeddings can find related deep dives on daily.dev._ ## Similar posts on daily.dev - [Nemotron ColEmbed V2: Raising the Bar for Multimodal Retrieval with ViDoRe V3’s Top Model](https://daily.dev/posts/nemotron-colembed-v2-raising-the-bar-for-multimodal-retrieval-with-vidore-v3-s-top-model-9yhpl9gtr) · Hugging Face · 0 upvotes · 0 comments - [Nvidia’s best model is now live](https://daily.dev/posts/nvidia-s-best-model-is-now-live-tsagrzqp8) · The New Stack · 0 upvotes · 0 comments - [NVIDIA Nemotron 3 Ultra is Here – and it’s Free to Use in Kilo](https://daily.dev/posts/nvidia-nemotron-3-ultra-is-here-and-it-s-free-to-use-in-kilo-qmednaxh7) · Kilo Blog · 1 upvotes · 0 comments --- Tags: [#ai-agents](https://daily.dev/tags/ai-agents), [#nlp](https://daily.dev/tags/nlp), [#nvidia](https://daily.dev/tags/nvidia), [#rag](https://daily.dev/tags/rag), [#embeddings](https://daily.dev/tags/embeddings) [View this post on daily.dev](https://daily.dev/posts/nvidia-nemotron-3-embed-8b-takes-1-on-rteb-multilingual-retrieval-leaderboard-b6jrl3kp7) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"NVIDIA Nemotron 3 Embed 8B takes #1 on RTEB multilingual retrieval leaderboard","url":"https://daily.dev/posts/nvidia-nemotron-3-embed-8b-takes-1-on-rteb-multilingual-retrieval-leaderboard-b6jrl3kp7","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/nvidia-nemotron-3-embed-8b-takes-1-on-rteb-multilingual-retrieval-leaderboard-b6jrl3kp7"},"datePublished":"2026-07-16T16:27:09.940Z","dateModified":"2026-09-13T19:15:28.848Z","description":"NVIDIA released Nemotron 3 Embed, a family of three open embedding models (8B and two 1B variants), with the 8B model claiming the top spot on the RTEB...","isAccessibleForFree":true,"articleSection":"Collections","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Collections","logo":"https://media.daily.dev/image/upload/s--fk_6ycEi--/f_auto,q_auto/v1780996001/logos/collections?_a=BAMAMiWQ0","url":"https://daily.dev/sources/collections"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/nvidia-nemotron-3-embed-8b-takes-1-on-rteb-multilingual-retrieval-leaderboard-b6jrl3kp7","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"ai-agents,nlp,nvidia,rag,embeddings","timeRequired":"PT2M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Collections","item":"https://daily.dev/sources/collections"},{"@type":"ListItem","position":3,"name":"NVIDIA Nemotron 3 Embed 8B takes #1 on RTEB multilingual retrieval leaderboard"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/nvidia-nemotron-3-embed-8b-takes-1-on-rteb-multilingual-retrieval-leaderboard-b6jrl3kp7#faq","mainEntity":[{"@type":"Question","name":"What score does NVIDIA Nemotron 3 Embed 8B get on the RTEB retrieval leaderboard?","acceptedAnswer":{"@type":"Answer","text":"Nemotron 3 Embed 8B (BF16) scores 78.5% on the RTEB multilingual retrieval leaderboard, taking the top spot. It supports a 32K context window and multilingual text and code retrieval, and is built on a Ministral-3B-8B-Instruct backbone converted into a bidirectional encoder trained with a contrastive objective. Teams choosing an embedding model for RAG can track daily.dev for how retrieval benchmarks like this evolve."}},{"@type":"Question","name":"What is the difference between Nemotron 3 Embed 1B BF16 and 1B NVFP4 variants?","acceptedAnswer":{"@type":"Answer","text":"The 1B BF16 variant scores 72.4% on RTEB, a 27% error-rate reduction over its predecessor, while the 1B NVFP4 variant is FP4-quantized specifically for NVIDIA Blackwell GPUs, delivering up to 2x the throughput of the BF16 version while retaining over 99% of its accuracy. Both target latency-sensitive deployments. Engineers weighing latency versus accuracy tradeoffs in embedding models can follow this comparison on daily.dev."}},{"@type":"Question","name":"How were the smaller Nemotron 3 Embed 1B models trained from a larger model?","acceptedAnswer":{"@type":"Answer","text":"NVIDIA first applied a bidirectional encoder conversion recipe to a 3B model, then used structured pruning and neural architecture search to shrink it to 2B, then distilled against the 8B model to recover quality. This entire pruning-and-distillation cycle was repeated a second time to produce the final 1B model. Practitioners exploring model compression techniques for embeddings can find related deep dives on daily.dev."}}]} ```