NVIDIA Nemotron 3 Embed 在 RTEB 综合排名第一，推动智能体检索技术发展

检索在多步骤智能体工作流中至关重要，糟糕的检索可能导致智能体抓取无关上下文、重复查询、浪费 token 预算，并将噪声带入后续推理步骤。

今天，我们发布 NVIDIA Nemotron 3 Embed，一系列开放且可商用的嵌入模型，旨在提升检索质量，同时为开发者提供实际的生产级 RAG、智能体检索、代码检索和智能体记忆部署选择。

该系列包含三款开放模型，在精度-效率曲线上均实现了最先进的检索性能，其中包括在 RTEB 排行榜上拔得头筹的 8B 模型，以及专为生产级部署构建的高效 1B 变体：

模型	角色	最适合
Nemotron-3-Embed-8B-BF16	旗舰质量锚点：旗舰嵌入模型，RTEB 排名第一。	精度关键的检索和高风险企业 RAG
Nemotron-3-Embed-1B-BF16	高效标准版：适用于延迟和成本敏感的生产检索的高效模型。	对成本和延迟敏感的生产服务
Nemotron-3-Embed-1B-NVFP4	硬件加速变体：面向 Blackwell 架构优化的变体，用于高吞吐量检索，内存占用更小。	超高吞吐量和大规模基础设施

表 1. Nemotron 3 Embed 模型可用性与部署矩阵。

图 1. RTEB 多语言排行榜 截图（2026 年 7 月 15 日），显示 Nemotron-3-Embed-8B-BF16 排名第 1。

主要特性

除了 RTEB 的结果外，Nemotron 3 Embed 还为企业检索部署引入了一系列生产就绪的功能：

• 开放权重、数据集和配方： 使团队能够自行检查、调整、微调并在自有基础设施上部署检索模型。
• 32k 上下文窗口： 支持对长文档、大代码上下文和多轮智能体历史记录进行检索，同时减少截断。
• 多语言与代码检索： 支持跨全球企业数据、技术文档和多文件代码仓库的检索。
• NVIDIA NVFP4 效率： 提供面向 Blackwell 优化的 4-bit 部署路径，用于高吞吐量检索，内存占用更小。
• 微调 和 蒸馏 配方： NVIDIA NeMo AutoModel 配方支持领域适应和模型压缩，适合希望将检索模型适配到自有数据的团队。
• 即日生态系统集成： 立即在 Hugging Face 上可用，可作为 NVIDIA NIM 微服务部署，得到 vLLM 支持，并可通过领先的 AI 云和推理合作伙伴访问。

评估：检索质量、智能体效率与部署权衡

我们从三个维度评估 Nemotron 3 Embed：检索质量、下游智能体效率和部署权衡。8B 模型确立了该模型系列的质量上限，而 1B BF16 和 NVFP4 变体则将相同的以检索为中心的设计带到了更低成本和更高吞吐量的部署场景中。

RTEB 领先并在检索基准上实现大幅提升

我们首先在 RTEB 上评估了模型，Nemotron-3-Embed-8B-BF16 排名第 1。我们还使用平均 NDCG@10 在 ViDoRe V3 Text、MMTEB Retrieval 和 LongEmbed 上对这些模型进行了测试。

图 2. 在 RTEB、ViDoRe V3 Text、MMTEB Retrieval 和 LongEmbed 上使用平均 NDCG@10 的检索准确性，对比 Nemotron 3 Embed 模型与上一代 Nemotron 基线。

• Nemotron-3-Embed-8B-BF16 在 RTEB 上排名第 1，RTEB 得分为 78.5%，MMTEB Retrieval 得分为 75.5%。
• Nemotron-3-Embed-1B-BF16 将 8B 模型的大量检索质量压缩进更小的部署占用空间。它在 RTEB 上得分为 72.4%，相对于其 1B 前身 (llama-nemotron-embed-vl-1b-v2) 将错误率降低了 27%，在 MMTEB Retrieval 上得分为 71.0%，将错误率降低了 28%。

为什么更好的检索对智能体至关重要

为了在智能体环境中评估检索，我们使用由 Nemotron 3 Ultra 驱动的搜索智能体，并改变检索系统使用的嵌入模型。更好的检索可以更早地返回相关证据，帮助智能体避免重复搜索、不必要的推理轮次和额外的上下文检查。我们比较了 ViDoRe V3、BRIGHT 和 BrowseComp-Plus 上的平均检索准确度与每次查询的估计下游智能体 token 成本。

图 3. 在 ViDoRe V3、BRIGHT 和 BrowseComp-Plus 上，平均检索准确度与每次查询的下游智能体 token 成本的关系。

评估说明：搜索智能体使用 Nemotron 3 Ultra。下游 token 成本根据 Nemotron 3 Ultra 的输入/输出 token 数量，使用 GPT-5.5 定价公式估算得出。

图 3 显示，更强的检索能力可以降低下游智能体的 token 成本。更准确的检索器能更早地返回相关证据，从而帮助智能体用更少的重复搜索和推理轮次完成任务。在这些评估中，Nemotron 3 Embed 模型改善了智能体检索的前沿水平，8B 模型在 ViDoRe V3、BRIGHT 和 BrowseComp-Plus 上同时提供了最高的平均检索准确度和最低的估计下游 token 成本。

在 Blackwell 上使用 NVFP4 扩展检索

对于高吞吐量部署，团队通常选择较小的嵌入模型来满足延迟和成本目标。Nemotron-3-Embed-1B-NVFP4 旨在通过利用 NVIDIA Blackwell 架构上的原生 NVFP4 加速，缩小服务效率与检索质量之间的差距。该模型将线性层的权重和激活量化为 NVFP4 以实现高效推理，并使用量化感知蒸馏 (QAD) 来帮助恢复长输入序列的准确性。

图 4. ViDoRe V3 检索准确度与服务效率对比，比较 Nemotron-3-Embed-1B-NVFP4 与精选的较小开源嵌入基线，包括 Qwen3-Embedding-0.6B 和 EmbeddingGemma-300M。

• 服务效率：在 Blackwell 上，NVFP4 相比 BF16 提供高达 2 倍的吞吐量，用于高吞吐量、低延迟的检索服务。
• 准确度保持：NVFP4 变体在减少内存占用的同时，保持了 99% 以上的 BF16 检索准确度。

即日即可使用的性能卓越的 NIM

对于生产级检索系统，服务栈还需要在实际请求负载、不同输入序列长度和硬件目标下保持这种效率。为使 Nemotron 3 Embed 从今日起便在企业规模上表现出色，我们还为 1B 模型发布了一个优化版的 NVIDIA NIM 微服务。如图 5 所示，基于 Rust 的 Nemotron 3 Embed NIM 在 NVIDIA GB200 和 RTX PRO 6000 GPU 上，在输入序列长度为 256 和 1024 时，性能与 vLLM 检查点相当或更优。

图 5. Nemotron 3 Embed NIM 服务性能与 NVIDIA GB200 和 RTX PRO 6000 GPU 上的 vLLM 检查点对比。

我们如何构建 Nemotron 3 Embed 模型

Nemotron-3-Embed-8B-BF16 适配了 Ministral-3-8B-Instruct-2512 核心架构，将其因果解码器转换为双向编码器，以进行全序列检索。该模型首先在网络来源和合成文本对混合数据上进行对比预训练，然后在涵盖法律、金融、医疗、商业和教育等领域的精选多语言检索数据集上进行微调。此 8B 模型作为旗舰嵌入模型，而同一开发线上的早期 8B 教师检查点则被用于蒸馏高效的 1B 变体。

缩减至 1B

1B 模型并非从头训练的小型检索器。我们首先将双向适应配方应用于 Ministral-3-3B-Instruct-2512 核心架构，建立了一个 3B 检索器基础，然后通过两轮结构化剪枝和蒸馏将其压缩。

首先，使用 NVIDIA ModelOpt 的 mcore_minitron 神经架构搜索引擎，将 3B 父模型压缩到 2B 中间规模。NAS 流水线在严格的参数量预算下搜索隐藏宽度、FFN 大小、注意力头数和深度，以识别适用于检索工作负载的高效架构。

然后，借助 8B 教师检查点对得到的 2B 中间模型进行蒸馏，以恢复排序准确度。我们使用了余弦距离损失和均方误差损失的组合，在多语言、领域内检索数据混合上，将学生的嵌入与教师对齐。

图 6. 剪枝与蒸馏流水线将检索器从 3B 基础模型压缩至最终的 1B 生产模型。

同样的顺序，即 ModelOpt 结构化剪枝后再进行 8B 教师蒸馏，被重复执行第二次，以将 2B 中间模型压缩至最终的 1.14B 嵌入模型。最终训练采用了一个渐进式两阶段上下文长度扩展计划：

• 阶段 1：专注于在 1024 token 上下文长度下的广泛多语言对齐，以重建父模型的核心检索行为。
• 阶段 2：将上下文长度扩展到 4096 token，并增加长上下文合成与推理数据集，帮助 1B 模型在更长的输入上保持辨别召回能力。

下表总结了 Nemotron 3 Embed 模型的核心技术规格和部署目标：

模型	规模	嵌入维度	上下文窗口	池化	输入前缀	目标硬件
Nemotron-3-Embed-8B-BF16	8.0B	4096	32k	均值	query: / document:	通用 GPU 推理
Nemotron-3-Embed-1B-BF16	1.14B	2048	32k	均值	query: / document:	低延迟 CPU/GPU
Nemotron-3-Embed-1B-NVFP4	1.14B	2048	32k	均值	query: / document:	NVIDIA Blackwell/GB200

表 2. Nemotron 3 Embed 模型的架构规格与核心推理配置。

企业合作伙伴评估

企业级独立软件供应商、AI 原生公司和内存提供商已经开始在智能体检索、智能体记忆、代码检索和生产推理工作流中评估 Nemotron 3 Embed。

• "上下文是智能体准确性的关键。我们的上下文智能图使用嵌入和语义相似性，向如 EnterpriseClaw 等智能体提供最相关的企业上下文——我们于五月与 NVIDIA 共同发布了该智能体。NVIDIA 全新 Nemotron 3 Embed 模型的早期结果令人鼓舞，尤其是在问答方面，与当前模型相比表现有所提升。我们对其进一步提升我们企业智能体准确性与可靠性的潜力感到兴奋。" - Adi Kuruganti，Automation Anywhere 首席 AI 与开发官
• "我们对 NVIDIA Nemotron 3 Embed 模型的初步评估显示，在我们的智能体检索用例中表现出强大的检索性能。1B 和 8B 变体的可用性为团队提供了在不同环境中平衡质量、延迟和部署需求的灵活性。我们期待继续评估这些模型，并探索它们如何支持用于生产 AI 应用的高性能检索。" - Mani Gill，Boomi 产品管理高级副总裁
• IBM 在评估新的 NVI...

NVIDIA Nemotron 3 Embed Ranks #1 Overall on RTEB, Advancing Agentic Retrieval

Retrieval is critical in multi-step agentic workflows where poor retrieval can cause agents to fetch irrelevant context, re-query, waste token budget, and carry noise into later reasoning steps.

Today, we are releasing NVIDIA Nemotron 3 Embed, a collection of open and commercially available embedding models designed to improve retrieval quality while giving developers practical deployment options for production-scale RAG, agentic retrieval, code retrieval, and agent memory.

The collection includes three open models that achieve state-of-the-art retrieval across the accuracy-efficiency curve, led by an 8B model that tops the RTEB leaderboard and efficient 1B variants built for production-scale deployment:

Model	Role	Best for
Nemotron-3-Embed-8B-BF16	Flagship Quality Anchor: The flagship embedding model, ranking #1 on RTEB.	Precision-critical retrieval and high-stakes enterprise RAG
Nemotron-3-Embed-1B-BF16	High-Efficiency Standard: A high-efficiency model for production retrieval where latency and cost matter.	Cost- and latency-sensitive production serving
Nemotron-3-Embed-1B-NVFP4	Hardware-Accelerated Variant: A Blackwell-optimized variant for high-throughput retrieval with a smaller memory footprint.	Ultra-high-throughput and massive-scale infrastructure

Table 1. Nemotron 3 Embed Model Usability and Deployment Matrix.

Figure 1.RTEB Multilingual Leaderboard screenshot (July 15, 2026) showing Nemotron-3-Embed-8B-BF16 ranked as #1.

Key Features

Beyond the RTEB result, Nemotron 3 Embed introduces a production-ready feature set for enterprise retrieval deployments:

• Open Weights, Datasets, and Recipes: Gives teams control to inspect, tune, fine-tune, and deploy retrieval models on their own infrastructure.
• 32k Context Window: Supports retrieval over long documents, large code contexts, and multi-turn agent histories while reducing truncation.
• Multilingual & Code Retrieval: Supports retrieval across global enterprise data, technical documentation and multi-file code repositories.
• NVIDIA NVFP4 Efficiency: Provides a Blackwell-optimized 4-bit deployment path for high-throughput retrieval with a smaller memory footprint.
• Fine-Tuning and Distillation Recipes: NVIDIA NeMo AutoModel recipes support domain adaptation and model compression for teams adapting retrieval models to their own data.
• Day-0 Ecosystem Integration: Available immediately on Hugging Face, deployable as NVIDIA NIM microservice, supported by vLLM, and accessible through leading AI Cloud and inference partners.

Evaluation: Retrieval Quality, Agentic Efficiency, and Deployment Tradeoffs

We evaluate Nemotron 3 Embed across three dimensions: retrieval quality, downstream agentic efficiency, and deployment tradeoffs. The 8B model establishes the model collection’s quality ceiling, while the 1B BF16 and NVFP4 variants bring the same retrieval-focused design to lower-cost and higher-throughput deployment settings.

RTEB Leadership and Strong Gains Across Retrieval Benchmarks

We first evaluated the models on RTEB, where Nemotron-3-Embed-8B-BF16 ranks #1. We also tested these models across ViDoRe V3 Text, and MMTEB Retrieval and LongEmbed using average NDCG@10.

Figure 2. Retrieval accuracy using average NDCG@10 across RTEB, ViDoRe V3 Text, MMTEB Retrieval and LongEmbed, comparing the Nemotron 3 Embed models with prior-generation Nemotron baselines.

• Nemotron-3-Embed-8B-BF16 ranks #1 on RTEB, scoring 78.5% on RTEB and 75.5% on MMTEB Retrieval.
• Nemotron-3-Embed-1B-BF16 brings much of the 8B model’s retrieval quality into a smaller deployment footprint. It scores 72.4% on RTEB, reducing error rate by 27% over its 1B predecessor (llama-nemotron-embed-vl-1b-v2), and scores 71.0% on MMTEB Retrieval, reducing error rate by 28%.

Why Better Retrieval Matters for Agents

To evaluate retrieval in an agentic setting, we use a search agent powered by Nemotron 3 Ultra and vary the embedding model used by the retrieval system. Better retrieval can return relevant evidence earlier, helping the agent avoid repeated searches, unnecessary reasoning turns, and extra context inspection. We compare average retrieval accuracy with estimated downstream agentic token cost per query across ViDoRe V3, BRIGHT, and BrowseComp-Plus.

Figure 3. Average retrieval accuracy versus downstream agentic token cost per query across ViDoRe V3, BRIGHT, and BrowseComp-Plus.

Evaluation note: The search agent uses Nemotron 3 Ultra. Downstream token cost is estimated from Nemotron 3 Ultra input/output token counts using the GPT-5.5 pricing formula.

Figure 3 shows that stronger retrieval reduces downstream agentic token cost. More accurate retrievers return relevant evidence earlier, which helps agents complete tasks with fewer repeated searches and fewer reasoning turns. In these evaluations, the Nemotron 3 Embed models improve the agentic retrieval frontier, with the 8B model delivering both the highest average retrieval accuracy and the lowest estimated downstream token cost across ViDoRe V3, BRIGHT, and BrowseComp-Plus.

Scaling Retrieval with NVFP4 on Blackwell

For high-throughput deployments, teams often choose smaller embedding models to meet latency and cost targets. Nemotron-3-Embed-1B-NVFP4 is designed to narrow the gap between serving efficiency and retrieval quality by using native NVFP4 acceleration on NVIDIA Blackwell architectures. The model quantizes the weights and activations of linear layers to NVFP4 for efficient inference, and uses Quantization-Aware Distillation (QAD) to help recover accuracy for long input sequences.

Figure 4. ViDoRe V3 retrieval accuracy versus serving efficiency, comparing Nemotron-3-Embed-1B-NVFP4 with selected smaller open embedding baselines, including Qwen3-Embedding-0.6B and EmbeddingGemma-300M.

• Serving Efficiency : NVFP4 on Blackwell delivers up to 2x higher throughput than BF16 for high-throughput, low-latency retrieval serving.
• Accuracy retention : The NVFP4 variant retains 99%+ of BF16 retrieval accuracy while reducing memory footprint.

Day 0 Performant NIM

For production-scale retrieval systems, the serving stack also needs to preserve that efficiency under real request loads, across different input sequence lengths and hardware targets. To make Nemotron 3 Embed performant at enterprise scale today, we are also releasing an optimized NVIDIA NIM microservice for the 1B model. As shown in Figure 5, the Rust-based Nemotron 3 Embed NIM matches or outperforms the vLLM checkpoint on NVIDIA GB200 and RTX PRO 6000 GPUs across ISLs of 256 & 1024.

Figure 5. Nemotron 3 Embed NIM serving performance compared with the vLLM checkpoint on NVIDIA GB200 and RTX PRO 6000 GPUs.

How We Built the Nemotron 3 Embed Models

Nemotron-3-Embed-8B-BF16 adapts the Ministral-3-8B-Instruct-2512 backbone by converting its causal decoder into a bidirectional encoder for full-sequence retrieval. The model is trained with contrastive pre-training on a blend of web-sourced and synthetic text pairs, then fine-tuned on curated multilingual retrieval datasets across domains such as legal, finance, medical, business, and education. This 8B model serves as the flagship embedding model, while earlier 8B teacher checkpoints from the same development line were used to distill the efficient 1B variants.

Scaling Down to 1B

The 1B model is not a small retriever trained from scratch. We first applied the bidirectional adaptation recipe to the Ministral-3-3B-Instruct-2512 backbone to establish a 3B retriever base, then compressed it through two-rounds of structured pruning and distillation.

First, the 3B parent model was compressed to a 2B intermediate footprint using NVIDIA ModelOpt’s mcore_minitron Neural Architecture Search engine. The NAS pipeline searched across hidden width, FFN size, attention heads, and depth under a strict parameter budget to identify an efficient architecture for retrieval workloads.

The resulting 2B intermediate model was then distilled from an 8B teacher checkpoint to recover ranking accuracy. We used a combined cosine distance loss and mean squared error loss on a multilingual, in-domain retrieval data blend to align the student’s embeddings with the teacher.

Figure 6. The pruning and distillation pipeline compresses the retriever from a 3B base to the final 1B production model.

This same sequence, ModelOpt structured pruning followed by 8B teacher distillation, was repeated a second time to compress the 2B intermediate model down to the final 1.14B embedding model. Final training used a progressive two-stage context-scaling schedule:

• Stage 1: Focused on broad multilingual alignment at 1024-token context length to reconstruct the core retrieval behavior of the parent model.
• Stage 2: Expanded context length to 4096 tokens and added long-context synthetic and reasoning datasets, helping the 1B model retain discriminative recall across longer inputs.

The following table summarizes the core technical specifications and deployment targets for the Nemotron 3 Embed models:

Model	Size	Emb Dim	Context Window	Pooling	Input Prefix	Target Hardware
Nemotron-3-Embed-8B-BF16	8.0B	4096	32k	Mean	query: / document:	General GPU Inference
Nemotron-3-Embed-1B-BF16	1.14B	2048	32k	Mean	query: / document:	Low-latency CPU/GPU
Nemotron-3-Embed-1B-NVFP4	1.14B	2048	32k	Mean	query: / document:	NVIDIA Blackwell/GB200

Table 2. Architectural specifications and core inference configurations for the Nemotron 3 Embed models.

Enterprise Partner Evaluations

Enterprise ISVs, AI-native companies, and memory providers are already evaluating Nemotron 3 Embed across agentic retrieval, agent memory, code retrieval, and production inference workflows.

• "Context is the key to agentic accuracy. Our Context Intelligence Graph uses embeddings and semantic similarity to deliver the most relevant enterprise context to agents like EnterpriseClaw, which we launched with NVIDIA in May. Early results from NVIDIA’s new Nemotron 3 Embed models are promising, particularly for question answering, where they show improvements over our current model. We’re excited about their potential to further improve the accuracy and reliability of our enterprise agents." - Adi Kuruganti, Chief AI and Development Officer at Automation Anywhere
• "Our initial evaluation of the NVIDIA Nemotron 3 Embed models shows strong retrieval performance for our agentic retrieval use cases. The availability of both 1B and 8B variants gives teams the flexibility to balance quality, latency, and deployment requirements across different environments. We're excited to continue evaluating the models and exploring how they can support high-performance retrieval for production AI applications." - Mani Gill, Senior Vice President of Product Management at Boomi
• IBM has seen promising early results evaluating the new NVIDIA Nemotron Embed model in a proof-of-concept built on watsonx.data.
• "We ran Nemotron-3-Embed-1B inside our own stack against various models and it came out on top. For instance, it scored 80.38% against Qwen-3-0.6B’s 78.71% on LongMemEval for Retrieval@10. The open weights and fine-tuning recipes are what interest us most, since they let us adapt the model to memory text and serve it ourselves." - Rudraj Mehta, Product Manager at Mem0
• Palantir is collaborating with NVIDIA to evaluate the deployment of Nemotron 3 Embed for end customers conducting edge retrieval workloads, building on the prior work making NVIDIA embedding models available in AIP's Model Catalog.
• ServiceNow is evaluating NVIDIA Nemotron Embed to retrieve through their documentation, with a continued interest in in-domain fine-tuning to improve their retrieval accuracy.
• turbopuffer is bringing Nemotron 3 Embed to its native embeddings service, enabling developers to use the models within its semantic search engine.
• "Swapping in NVIDIA Nemotron 3 Embed model within our re-ranking stack delivered a significant leap in performance, letting us select query-relevant snippets from web pages with far greater accuracy than any model we'd used before." - Rahul Mohan, Senior AI Engineer, You.com
• Zep evaluated NVIDIA Nemotron 3 Embed 1B for agent memory and context retrieval. In early internal benchmarks against several models used by Zep, including some much larger than Nemotron 3 Embed 1B, the model ranked first across all memory retrieval tasks. Zep believes the availability of an FP4 checkpoint also makes the model very storage and latency competitive for their scale.
• Zoom is evaluating NVIDIA Nemotron 3 Embed for the retrieval layer behind the enterprise agentic search that powers Zoom's contextual layer, which helps agents retrieve relevant context across meetings and workplace knowledge sources, including docs, chat, tickets, and other internal systems.

Getting Started

NVIDIA Nemotron 3 Embed is released with open weights and open-source training recipes, giving organizations full control over how retrieval models are customized and deployed for production AI applications.

Developers can get started using the deployment option that best fits their workflow:

• Hugging Face - Access the model weights, model cards, and example code for SentenceTransformers, Transformers, and vLLM.
• NVIDIA NIM - Access optimized microservice on build.nvidia.com for production-ready inference.
• AI Cloud and Inference Partners - Deploy Nemotron 3 Embed through leading ecosystem partners - Baseten, Bitdeer AI, DeepInfra, Friendli AI, OpenRouter.

For workloads requiring domain adaptation or footprint reduction, we have also open-sourced NVIDIA NeMo AutoModel training recipes:

• Fine-tuning recipe - Adapt Nemotron 3 Embed to your enterprise corpus and retrieval tasks.
• Distillation recipe - Compress larger retrieval models while retaining ranking quality for production deployment.

For example, on the NV Docs evaluation, fine-tuning Nemotron-3-Embed-1B-BF16 improved NDCG@10 from 56.7% to 63.3% (+11.6%) and Recall@5 from 56.1% to 62.8% (+11.9%).

Whether you are building enterprise search, production RAG, agent memory, code retrieval, or agentic AI systems, NVIDIA Nemotron 3 Embed provides flexible deployment options—from open-source models on Hugging Face to fully managed AI Cloud platforms and NVIDIA NIM microservices.

Explore the models, deploy them on your preferred platform, and let us know what you are building.