NVIDIA Nemotron 3.5 Embed、Rerank 与 Parse：以世界一流的检索模型推动智能体发展

今天，我们将通过新增开放且可商用的嵌入、重排序和文档提取模型，扩展 Nemotron Retriever 系列。新增模型包括 Nemotron 3.5 Embed 8B BF16、1B BF16、1B NVFP4、Nemotron 3.5 Rerank 8B BF16，以及拥有 6.31 亿参数的 Nemotron 3.5 Parse。Today, we are expanding the Nemotron Retriever collection with new open, commercially available models for embedding, reranking, and document extraction. The new models include Nemotron 3.5 Embed 8B BF16, 1B BF16, 1B NVFP4, Nemotron 3.5 Rerank 8B BF16, and Nemotron 3.5 Parse at 631M parameters.

这些先进模型共同为企业智能体工作流带来了多模态检索、相关性排序和文档处理能力。Together, these state-of-the-art models bring multimodal retrieval, relevance ranking, and document processing to enterprise agent workflows.

更优检索，让智能体更高效 Better Retrieval for More Efficient Agents

检索质量会影响智能体生成的答案以及 LLM 推理成本。相关证据有助于智能体回答问题，或更好地判断下一步行动；重排序则能提升进入其上下文的证据质量。Retrieval quality affects both the answers an agent produces and the LLM inference cost. Relevant evidence can help the agent answer a question or better decide its next step, and reranking improves the quality of evidence that enters its context.

设想有个智能体被要求识别下个季度面临缺货风险的产品，并解释原因。它必须综合图表中的需求预测、表格中的库存水平以及报告中的供应商交货周期，然后判断补货能否在库存耗尽前到货。要完成这项任务，就需要从各类文档中找到相关证据，并在分析中综合运用。Consider an agent asked to identify products at risk of stockouts next quarter and explain why. It must combine demand forecasts in charts, inventory levels in tables, and supplier lead times in reports, then reason about whether replenishment will arrive before inventory runs out. Completing the task requires finding the relevant evidence across documents and using it together in the analysis.

为了了解检索如何影响智能体整合证据并完成任务的能力，我们在 ViDoRe v3、BRIGHT 和 BrowseComp-Plus 上评估了 Nemotron 3.5 Embed 和 Rerank。ViDoRe v3 使用文本与图像组合输入，而 BRIGHT 和 BrowseComp-Plus 仅使用文本输入。在此流程中，智能体通过搜索工具查询语料库、检查证据，并优化下一次搜索尝试，以找到与给定查询最相关的信息。以下实验改变了搜索工具的实现方式：仅使用嵌入、使用嵌入加重排，以及为两者分别使用不同模型。智能体所用的 LLM、工具、任务集、语料库和执行策略均保持不变，仅调整检索配置。我们衡量智能体为每项任务找到相关信息的能力，以及每项任务的智能体推理总 token 数，其中包括所有轮次的输入和输出 token。To understand how retrieval affects the agent’s ability to combine evidence and complete a task, we evaluated Nemotron 3.5 Embed and Rerank across ViDoRe v3, BRIGHT, and BrowseComp-Plus. ViDoRe v3 uses combined text and image inputs, while BRIGHT and BrowseComp-Plus use text-only inputs. In this pipeline, an agent queries the corpus using a search tool, inspects the evidence, and refines its next search attempt to find the most relevant information to a given query. The experiments below vary how the search tool is implemented: embedding only, embedding plus reranking, and with different models for each. The LLM powering the agent, tools, task set, corpus, and execution policy remain fixed while the retrieval configuration changes. We measure the agent's ability to find the relevant information for each task and total agent inference tokens per task, including input and output tokens across all turns.

图 1. 不同检索配置下，端到端任务成功率与每项任务的智能体推理令牌总数之间的关系。Figure 1. End-to-end task success versus total agent inference tokens per task across retrieval configurations.

• 降低智能体推理成本：使用 Nemotron 3.5 Embed 8B 时，添加 Nemotron 3.5 Rerank 可将每次查询的智能体令牌成本降低约 14%；使用 Nemotron 3.5 Embed 1B 时则可降低约 13%，同时提升这两种配置下的检索准确率。
  Lower agent inference cost: Adding Nemotron 3.5 Rerank reduces agentic token cost per query by approximately 14% with Nemotron 3.5 Embed 8B, and 13% with Nemotron 3.5 Embed 1B, while improving retrieval accuracy in both configurations.
• 在相同的智能体预算下获得更优证据：以 Nemotron 3.5 Embed 8B 为检索器时，Nemotron 3.5 Rerank 的 nDCG@10 比 Qwen3-VL-Reranker-8B 高约 0.9 个百分点，而智能体 token 成本相近。
  Better evidence within the same agent budget: With Nemotron 3.5 Embed 8B as the retriever, Nemotron 3.5 Rerank delivered approximately 0.9 points higher nDCG@10 than Qwen3-VL-Reranker-8B at a similar agentic token cost.

支撑这些智能体取得成功的，是 Nemotron 3.5 Embed 和 Nemotron 3.5 Rerank：它们能从文本和图像中检索证据，并为智能体的上下文筛选优先级。Nemotron 3.5 Embed 基于上一代 Nemotron 3 Embed 文本嵌入模型构建，可从文档图像中检索证据，包括表格、图表、信息图、数据表和示意图。这使应用能够搜索视觉内容所传达的信息，以及文档文本中的信息。Behind these agent successes are Nemotron 3.5 Embed and Nemotron 3.5 Rerank, which retrieve evidence across text and images and prioritize it for the agent’s context. Nemotron 3.5 Embed builds on the previous generation’s Nemotron 3 Embed text embedding models to retrieve evidence from document images, including tables, charts, infographics, datasheets, and schematics. This allows applications to search information conveyed by visual content alongside document text.

这四种模型为检索和第二阶段排序提供了不同的选择：The four models provide different options for retrieval and second-stage ranking:

表 1. 不同规模和变体的 Nemotron 3.5 Embed 与 Nemotron 3.5 Rerank 模型的能力和应用场景。Table 1. Nemotron 3.5 Embed and Nemotron 3.5 Rerank model capabilities and use cases across various sizes and variants.

主要功能包括：Key features include:

• 文本和图像检索：搜索文档文本以及包含表格、图表、信息图、数据表和示意图的图像。
  Text and image retrieval: Search document text and images containing tables, charts, infographics, datasheets, and schematics.
• 多语言支持：文本检索在 34 种语言中进行了评估，多语言视觉检索则在 MIRACL Vision（18 种语言）和 KoViDoRe 上进行了评估。该模型可能还支持评估范围之外的其他语言。
  Multilingual support: Text retrieval was evaluated across 34 languages, while multilingual visual retrieval was evaluated on MIRACL Vision (18 languages) and KoViDoRe. The model may support additional languages beyond those evaluated.
• 多模态重排序：一个 8B 交叉编码器对文本和图像输入中的查询—文档对进行评分。
  Multimodal reranking: An 8B cross-encoder scores query–document pairs across text and image inputs.
• BF16 和 NVFP4 部署：该 1B 嵌入模型提供 BF16 和 NVFP4 版本，并对文本和图像编码器均进行了量化。
  BF16 and NVFP4 deployment: The 1B embedding model is available in BF16 and NVFP4, with quantization applied to both text and image encoders.
• 领域适配：Nemotron 嵌入模型微调方案支持自定义文档集合；在我们的图表和信息图实验中，Nemotron-3.5-Embed-1B-BF16 和 Nemotron-3.5-Embed-8B-BF16 的 nDCG@10 分别相对提升 18% 和 3%。
  Domain adaptation: Nemotron embedding fine-tuning recipes support custom document collections, with a 18% and 3% relative nDCG@10 gain for Nemotron-3.5-Embed-1B-BF16 and Nemotron-3.5-Embed-8B-BF16 respectively in our chart and infographic experiments.

在 ViDoRe v3 上领先的多模态稠密检索 Leading Multimodal Dense Retrieval on ViDoRe v3

在 ViDoRe v3.1 上，Nemotron 3.5 Embed 8B BF16 的 nDCG@10 达到 61.17，在稠密检索模型中排名第一。相比之下，Llama Nemotron Embed VL 1B 模型的得分为 52.23，前者高出 8.94 分。1B 模型将大部分这类视觉检索能力浓缩到了更小的模型中。Nemotron 3.5 Embed 1B BF16 由 8B 教师模型蒸馏而来，nDCG@10 得分为 58.38，保留了教师模型 95.4% 的得分。On ViDoRe v3.1, Nemotron 3.5 Embed 8B BF16 achieves 61.17 nDCG@10, ranking first among dense retrieval models. This is an 8.94-point improvement over the 52.23 score achieved by the Llama Nemotron Embed VL 1B model. The 1B model brings most of this visual retrieval quality into a smaller model. Distilled from the 8B teacher, Nemotron 3.5 Embed 1B BF16 scores 58.38 nDCG@10, retaining 95.4% of the teacher’s score.

图 2. 截至 2026 年 10 月 20 日，Nemotron 3.5 Embed 8B BF16 在 ViDoRe v3.1 排行榜的稠密检索模型中排名第一。Figure 2. Nemotron 3.5 Embed 8B BF16 ranked first among dense retrieval models on the ViDoRe v3.1 leaderboard as of 10/20/2026.

我们还在 ViDoRe v3、韩语 ViDoRe 和 MIRACL Vision 基准测试中，将这些模型与 Qwen3-VL-Embedding、EmbeddingGemma-2 以及 NVIDIA 上一代模型进行了比较。下图展示了仅输入图像以及同时输入图像和文本时的检索准确率。We also compare the models with Qwen3-VL-Embedding, EmbeddingGemma-2, and the previous-generation NVIDIA model across ViDoRe v3, Korean ViDoRe and MIRACL Vision. The chart below reports retrieval accuracy for image-only and combined image-and-text inputs.

图 3. Nemotron 3.5 Embed 与上一代模型及选定基线模型在文本和视觉检索方面的性能对比。Figure 3. Text and visual retrieval performance of Nemotron 3.5 Embed compared with prior-generation models and selected baselines.

多语言结果展示了这些模型在不同语言环境下检索证据的表现。在 MIRACL Vision 上，8B 模型的 nDCG@10 达到 72.6，比 Qwen3-VL-Embedding-8B 高 4.3 分。我们还在 KoViDoRe 上评估了这些模型；使用图像和文本组合输入时，Nemotron 3.5 Embed 8B 和 1B 的表现均优于 Qwen 基线模型。The multilingual results show how the models retrieve evidence across different language settings. On MIRACL Vision, the 8B model achieves 72.6 nDCG@10, exceeding Qwen3-VL-Embedding-8B by 4.3 points. We also evaluate the models on KoViDoRe; both Nemotron 3.5 Embed 8B and 1B outperform the Qwen baselines when using combined image and text inputs.

对智能体任务的证据重新排序 Reranking the Evidence for Agent Tasks

嵌入模型通过将内容表示为向量，让大型集合变得可搜索。检索返回候选集后，Nemotron 3.5 Rerank 8B BF16 会根据查询评估每个候选项，从而对文本和图像输入的排序进行优化。通过区分广泛相关的内容与能直接回答问题的证据，重排序模型有助于智能体更有效地利用上下文预算，并聚焦于完成任务所需的信息。对于智能体检索系统，更早呈现最有力的证据可以减少干扰性上下文，并帮助智能体生成依据更充分、效率更高的回答。Embedding models make large collections searchable by representing content as vectors. After retrieval returns a candidate set, Nemotron 3.5 Rerank 8B BF16 evaluates each candidate against the query to refine the ordering across text and image inputs. By distinguishing broadly related content from evidence that directly answers the question, the reranker helps agents use their context budget more effectively and focus on information needed to complete the task. For agentic retrieval systems, surfacing the strongest evidence earlier can reduce distracting context and help agents produce more grounded, efficient responses.

我们与 MTEB 团队合作推出 ViDoRe v3.1 重排序排行榜，用于评测重排序模型在仅图像、仅文本以及图文结合输入下的文档检索表现。Nemotron 3.5 Rerank 8B BF16 在图文结合输入下取得 60.49 的 nDCG@10，高于 Qwen-3-VL-Reranker-8B 的 59.17。In collaboration with the MTEB team, we are launching the ViDoRe v3.1 reranking leaderboard to benchmark rerankers on document retrieval across image-only, text-only, and combined image-and-text inputs. Nemotron 3.5 Rerank 8B BF16 achieves 60.49 nDCG@10 with combined image-and-text inputs compared with Qwen-3-VL-Reranker-8B’s 59.17.

图 4. 截至 2026 年 10 月 20 日，Nemotron-3.5-Rerank-8B-BF16 在 ViDoRe v3.1 重排序排行榜上位列第一。Figure 4. Nemotron-3.5-Rerank-8B-BF16 ranked first on ViDoRe v3.1 Reranking leaderboard as of 10/20/2026.

图 5. Nemotron 3.5 Rerank 8B 与 Qwen 3 VL Reranker 8B 的文本和视觉重排性能。两种模型均对 Nemotron 3.5 Embed 1B 模型检索出的前 50 个候选项进行了重排。Figure 5. Text and visual reranker performance of Nemotron 3.5 Rerank 8B compared with Qwen 3 VL Reranker 8B. Both models reranked the top 50 candidates retrieved by the Nemotron 3.5 Embed 1B model.

借助 NVFP4 扩展多模态检索 Scaling Multimodal Retrieval with NVFP4

对于生产环境中的智能体系统而言，检索质量只是其中一个考量因素。随着文档集合和查询量不断增长，团队必须在准确性、吞吐量和基础设施需求之间取得平衡。Retrieval quality is only part of the equation for production agentic systems. As document collections and query volumes grow, teams must balance accuracy, throughput, and infrastructure requirements.

Nemotron 3.5 Embed 1B NVFP4 为高吞吐量多模态检索提供了优化的部署选项。在 NVIDIA GB200 上，与 1B BF16 模型相比，其吞吐量提高了 27.1%，从每秒 101.3 篇文档提升至每秒 128.8 篇文档，同时在 ViDoRe v3 的图像与文本组合输入上保留了 99.85% 的检索准确率。如图 6 所示，对于需要在生产工作负载中扩展检索规模的团队而言，NVFP4 模型在准确率与吞吐量之间实现了出色的平衡。Nemotron 3.5 Embed 1B NVFP4 provides an optimized deployment option for high-throughput multimodal retrieval. On NVIDIA GB200, it delivers 27.1% higher throughput compared with the 1B BF16 model, increasing from 101.3 to 128.8 documents per second while retaining 99.85% of its retrieval accuracy on ViDoRe v3 with combined image and text inputs. As Figure 6 shows, the NVFP4 model offers a strong accuracy-throughput tradeoff for teams scaling retrieval across production workloads.

这种部署灵活性与领域定制相辅相成：团队可以根据自身基础设施选择合适的模型权重数据类型，然后针对自身的术语、文档和检索任务调整嵌入模型和重排序模型。This deployment flexibility complements domain-specific customization: teams can select the data type of model weights that fits their infrastructure, then adapt the embedding and reranking models to their own terminology, documents, and retrieval tasks.

图 6. ViDoRe V3 检索准确率与服务效率对比：Nemotron 3.5 Embed 与选定的小型开源嵌入基线模型。Figure 6. ViDoRe V3 retrieval accuracy versus serving efficiency, comparing Nemotron 3.5 Embed with selected smaller open embedding baselines.

使用 NVIDIA NIM 实现高性能服务部署 High-Performance Serving with NVIDIA NIM

NVIDIA NIM 为 Nemotron 3.5 Embed 提供了高性能部署方案。在 NVIDIA RTX PRO 6000 Blackwell Server Edition 上的测试中，NIM 的查询吞吐量达到每秒 6,031 个请求，约为可比 vLLM 配置（每秒 2,099 个请求）的 2.9 倍，同时保持了相当的段落处理吞吐量。这套优化的服务栈可帮助企业借助适用于生产环境的部署和支持，扩展检索工作负载。NVIDIA NIM provides a high-performance deployment path for Nemotron 3.5 Embed. In testing on an NVIDIA RTX PRO 6000 Blackwell Server Edition, NIM delivered 6,031 query requests per second—approximately 2.9× the throughput of a comparable vLLM configuration at 2,099 requests per second—while maintaining comparable passage-processing throughput. This optimized serving stack helps enterprises scale retrieval workloads with production-ready deployment and support.

Figure 7. Nemotron 3.5 Embed NIM serving performance compared with the vLLM checkpoint on NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs.

让检索适应特定领域的任务 Adapting Retrieval to Domain-Specific Tasks

企业级智能体经常需要搜索包含专业术语和公司特定信息的文档，而通用检索模型可能无法充分捕捉这些内容。微调使团队能够利用自己的文档、查询和相关性标签，根据特定领域和用例调整 Nemotron 3.5 Embed 和 Rerank。Enterprise agents often search documents containing specialized terminology and company-specific information that general-purpose retrieval models may not fully capture. Fine-tuning allows teams to adapt Nemotron 3.5 Embed and Rerank using their own documents, queries, and relevance labels to their specific domains and use cases.

Nemotron 多模态嵌入微调方案将合成数据生成、困难负样本挖掘、使用 NVIDIA NeMo AutoModel 进行微调，以及评估整合到一个工作流中。团队可以根据自己的文档生成查询和相关性标签，为检索任务构建训练数据。The Nemotron multimodal embedding fine-tuning recipe combines synthetic data generation, hard-negative mining, fine-tuning with NVIDIA NeMo AutoModel, and evaluation in one workflow. Teams can generate queries and relevance labels from their own documents to build training data for their retrieval tasks.

我们使用 Qwen3.8-Flash-Next 生成的 8,591 条合成查询，在 8,202 张图表和信息图图像上评估了该方案，重点关注涉及多张相关图像的枚举式和开放式查询。使用 6,872 条查询对 Nemotron-3.5-Embed-1B-BF16 微调一个 epoch 后，在 1,719 条留出的合成查询上，nDCG@10 从 37.34 提升至 44.12，相较于基础模型相对提升了 18%。使用同一组合成查询对 Nemotron-3.5-Embed-8B-BF16 微调一个 epoch 后，nDCG@10 从 41.44 提升至 42.75，相对提升了 3%。We evaluated the recipe on 8,202 chart and infographic images using 8,591 synthetic queries generated with Qwen3.8-Flash-Next, focusing on enumerative and open-ended queries with multiple relevant images. Fine-tuning the Nemotron-3.5-Embed-1B-BF16 on 6,872 queries for one epoch increased nDCG@10 from 37.34 to 44.12, a 18% relative gain over the base model, on 1,719 held-out synthetic queries. Using the same synthetic queries set, fine-tuning Nemotron-3.5-Embed-8B-BF16 for epoch increased nDCG@10 from 41.44 to 42.75, a 3% relative gain.

我们如何构建嵌入和重排序模型 How We Built the Embed and Rerank Models

Nemotron 3.5 Embed 8B 基于 Ministral-3-8B-Instruct 构建，其语言组件由 Nemotron-3-Embed-8B-BF16 初始化。该模型使用对比学习在图像和文本语料上进行训练，以生成用于多模态检索的表征。Nemotron 3.5 Embed 8B builds on Ministral-3-8B-Instruct, with its language component initialized from Nemotron-3-Embed-8B-BF16. The model is trained on image and text corpora using contrastive learning to produce representations for multimodal retrieval.

Embed 1B 模型由 Nemotron 3.5 Embed 8B 蒸馏而来，其语言组件以 Nemotron-3-Embed-1B-BF16 初始化。蒸馏过程结合余弦损失和均方误差（MSE）损失，使学生模型的嵌入与教师模型对齐。The Embed 1B models are distilled from Nemotron 3.5 Embed 8B, with their language component initialized from Nemotron-3-Embed-1B-BF16. Distillation combines cosine and mean squared error (MSE) losses to align the student’s embeddings with the teacher’s.

在嵌入模型训练的整个过程中，语言组件始终保持冻结。因此，文本嵌入预计将与相应的 Nemotron 3 Embed 模型保持一致，使现有文本嵌入和向量索引能够继续使用。The language component remains frozen throughout embedding-model training. Text-only embeddings are therefore expected to remain unchanged from the corresponding Nemotron 3 Embed models, allowing existing text embeddings and vector indexes to be reused.

Nemotron 3.5 Rerank 采用与 Nemotron-3.5-Embed-8B 相同的主干架构，并以交叉编码器配置运行。嵌入模型分别对查询和文档进行编码，并计算二者的相似度；而重排序模型则对每个查询—文档对进行联合编码。最终，多层感知机（MLP）评分头会输出一个标量相关性 logit。Nemotron 3.5 Rerank uses the same backbone architecture as Nemotron-3.5-Embed-8B in a cross-encoder configuration. While the embedding model encodes queries and documents separately and scores their similarity, the reranker jointly encodes each query–document pair. A final multilayer perceptron (MLP) scoring head produces a scalar relevance logit.

所有模型均使用公开可用数据和合成的多语言、多模态数据混合训练，最终发布的权重则通过检查点集成生成。有关训练数据的更多详情，请参阅各模型在 Hugging Face 上的模型卡片。All models are trained on a mixture of publicly available and synthetic multilingual, multimodal data, and their final released weights are produced through checkpoint ensembling. Further details on the training data are available in each model’s Hugging Face model card.

Nemotron 3.5 解析 Nemotron 3.5 Parse

企业知识通常以 PDF 的形式存在，而其中嵌入的文本只能反映页面内容的一部分。提取这些文本可能会打乱阅读顺序、使表格数值与表头分离，或遗漏图表所传达的信息。扫描文档和基于图像的 PDF 可能根本不包含可用的文本层。在这些信息能够用于搜索、检索或自动化文档工作流之前，需要先将像素内容转换为结构化内容。Enterprise knowledge often arrives in PDFs whose embedded text captures only part of what is on the page. Extracting that text can lose reading order, separate table values from their headers, or omit information conveyed in charts. Scanned documents and image-based PDFs may contain no usable text layer at all. Before this information can support search, retrieval, or automated document workflows, the pixels need to become structured content.

Nemotron 3.5 Parse 可将文档图像转换为机器可读文本，同时保留赋予内容意义的结构。它能够识别文档元素、元素位置及阅读顺序，并将表格和图表提取为结构化表示。生成的 Markdown、HTML 表格和结构化内容块，为开发者提供了更适合索引、搜索和后续处理的内容。应用可以检索带有上下文的段落，将表格作为行和列进行处理，并将提取出的证据关联回其在原始页面中的位置。Nemotron 3.5 Parse converts document images into machine-readable text while preserving the structure that gives the content meaning. It identifies document elements, their locations, and reading order, and extracts tables and charts into structured representations. The resulting Markdown, HTML tables, and structured blocks give developers content that is better suited to indexing, search, and downstream processing. Applications can retrieve a passage with its surrounding context, process a table as rows and columns, and link extracted evidence back to its location on the original page.

Nemotron 3.5 Parse 拥有 6.31 亿个参数，可将这一能力引入大规模文档处理流水线，并与 Nemotron 3.5 Embed 和 Rerank 相辅相成，为检索和智能体工作流提供结构化信息。At 631 million parameters, Nemotron 3.5 Parse brings this capability to document pipelines at scale, complementing Nemotron 3.5 Embed and Rerank with structured information for retrieval and agent workflows.

以更快的文档处理实现领先的准确性 Leading Accuracy with Faster Document Processing

图 8. 截至 2026 年 10 月 20 日，在参数量少于 10 亿的模型中，Nemotron 3.5 Parse 在 ParseBench 排行榜上排名第一。Figure 8. Nemotron 3.5 Parse ranked first on ParseBench leaderboard among models with fewer than 1 billion parameters as of 10/20/2026.

在参数量少于 10 亿的模型中，Nemotron 3.5 Parse 在 ParseBench 上排名第一，得分为 75.20，并且在单块 NVIDIA H100 上每秒可处理 10.5 页。它通过优化良好的参数分配、多 token 预测、推测解码以及灵活的原生分辨率处理，实现了同类最佳速度。ParseBench 测试表格和图表提取、内容忠实度、语义格式化以及视觉定位能力。Nemotron 3.5 Parse ranks #1 on ParseBench among models with fewer than 1 billion parameters, achieving a score of 75.20 while processing 10.5 pages per second on one NVIDIA H100. It delivers best-in-class speed with well optimized parameter allocation, multi-token prediction, speculative decoding and flexible native-resolution handling. ParseBench tests table and chart extraction, content faithfulness, semantic formatting, and visual grounding.

开发者可以使用同一模型，在处理速度与准确度之间进行权衡。单次推理只提取每个页面一次，以实现最高吞吐量。双次推理会重新处理初始页面解析中识别出的表格和图表。三次推理则会增加一个检测阶段，以查找更多需要细化处理的区域。这种方法会将额外的处理集中在值得仔细查看的文档区域。Developers can scale processing toward speed or accuracy using the same model. Single-pass inference extracts each page once for the highest throughput. Double-pass inference revisits tables and charts identified in the initial page parse. Triple-pass inference adds a detection pass to find additional regions for refinement. This approach concentrates additional processing on document regions that benefit from a closer look.

表 2. 模型准确率可通过增加模型处理轮次来提升，代价是吞吐量略有下降。对表格和图表的重点裁剪区域进行重新处理可提高准确率。Table 2. Model accuracy is able to scale with additional model passes at a small cost to throughput. Reprocessing focused crops of tables and charts leads to increased accuracy.

在这次比较中，这三种配置均为帕累托最优：没有其他参与比较的配置能同时实现更高的准确率和更高的吞吐量。团队可以根据工作负载选择合适的运行点，从快速摄取大量文档，到更细致地提取复杂页面中的信息。All three configurations are Pareto optimal within this comparison: no other compared configuration achieves both higher accuracy and higher throughput. Teams can choose the operating point that fits their workload, from rapid ingestion of large document collections to more detailed extraction of complex pages.

图 9. Nemotron Parse 3.5 在 ParseBench 排行榜上实现了帕累托最优性能。Figure 9. Nemotron Parse 3.5 achieves pareto optimal performance on ParseBench leaderboard.

How We Built Nemotron 3.5 Parse

Nemotron 3.5 Parse 使用多种企业文档、科学论文、表单、表格、图表和多语言文本混合数据进行训练。强大的合成数据引擎扩展了模型接触到的版式、结构和语言范围，并与自动标注文档和人工标注数据相辅相成。合成多语言数据有助于打造广泛的 OCR 能力，在涵盖 163 种语言与文字组合的评测中表现出色。这种多样性让模型能够处理企业文档集合中各种各样的内容。模型卡和多语言覆盖范围。Nemotron 3.5 Parse is trained on a diverse mixture of enterprise documents, scientific papers, forms, tables, charts, and multilingual text. A strong synthetic data engine expands the range of layouts, structures, and languages the model sees, complementing automatically labeled documents and human-labeled data. Synthetic multilingual data helps build broad OCR capability, with strong performance in evaluations spanning 163 language and script combinations. This diversity prepares the model for the varied content found in enterprise document collections. Model card and multilingual coverage.

我们进一步通过群组相对策略优化（GRPO）提升模型的稳健性和可靠性，并使用自定义奖励函数强化正确的文档结构、惩罚失败情况。这种后训练方法减少了重复生成循环和未闭合的内容块，提升了表格的一致性，并加强了对图表模式的遵循，使提取的信息在下游文档工作流中更加可靠。We further improve robustness and reliability through Group Relative Policy Optimization (GRPO), using custom reward functions to reinforce correct document structure and penalize failures. This post-training reduces repetitive generation loops and unclosed blocks, improves table consistency, and strengthens adherence to chart schemas, making extracted information more reliable for downstream document workflows.

感谢 LightOn 通过其模型和技术博客提升了开源领域的标准。他们分享的技术对改进我们的模型起到了关键作用。Thank you to LightOn for raising the open-source bar with their models and technical blog. The shared techniques were pivotal in improving our model.

早期企业合作伙伴评估 Early Enterprise Partner Evaluations

• Elastic 客户可以通过 NVIDIA 推理端点，使用 Elasticsearch 的 Inference API 调用 NVIDIA Nemotron Retriever 嵌入和重排序模型。Nemotron 3 即将通过 Elastic Inference Service (EIS) 提供，而对 Nemotron 模型本地部署的支持也已列入路线图。
  Elastic customers can access NVIDIA Nemotron Retriever embedding and reranking models through Elasticsearch’s Inference API using NVIDIA inference endpoints. Nemotron 3 will be available via Elastic Inference Service (EIS) soon, while support for Nemotron models on-prem deployment is on the roadmap.
• Vectorize 使用 Hindsight Agent Memory 对 NVIDIA Nemotron 3.5 Embed 进行了评估。该模型在两项长对话记忆基准测试中名列第一：在 LoCoMo 上的 recall@10 达到 93.8%，在 BEAM 100K 上达到 65.9%，击败了另外七种嵌入模型。它的 MRR 和 recall@5 也最高，凸显了其从漫长而复杂的对话历史中检索相关记忆的优势。
  Vectorize evaluated NVIDIA Nemotron 3.5 Embed using Hindsight Agent Memory, and it ranked first on two long-conversation memory benchmarks: achieving 93.8% recall@10 on LoCoMo and 65.9% on BEAM 100K against seven other embedding models. It also delivered the highest MRR and recall@5, highlighting its strength for retrieving relevant memories across long, complex conversation histories.
• Perplexity 在 Q2D-Web 基准测试中，NVIDIA Nemotron 3.5 Embed 表现强劲。在最新评测中，NVIDIA Nemotron 3.5 Rerank 8B 在包含 72 项任务的私有重排基准测试中取得了最高平均分，并在全部六个领域表现出色。这些结果凸显了专业检索模型与 Perplexity Decider 系列这类通用决策模型协同使用的价值。
  Perplexity saw strong results from NVIDIA Nemotron 3.5 Embed on the Q2D-Web benchmark. In their latest evaluation, NVIDIA Nemotron 3.5 Rerank 8B achieved the highest average score on the private 72-task reranking benchmark, with strong results across all six domains. These results highlight the value of specialized retrieval models alongside general-purpose decision models like Perplexity Decider family.
• Turbopuffer 在企业级 PDF 搜索基准测试中评估了 NVIDIA Nemotron 3.5 Embed 8B，发现与次优模型相比，其 Recall@20 指标下的相关结果多约 5%。这些模型在常见的业务任务中处于业界领先水平，例如搜索包含复杂图像的 PDF；现在也已在 Turbopuffer 原生嵌入功能中提供。
  Turbopuffer evaluated NVIDIA Nemotron 3.5 Embed 8B on the enterprise PDF search benchmark and found that it provided ~5% more relevant results at Recall@20 than the next best model. These models are SOTA for common business tasks, like searching through PDFs with complex images, and are now available in Turbopuffer native embedding.

快速入门 Getting Started

开发者可以在 Hugging Face 上获取开放权重、数据、训练配方等资源，部署 NVIDIA NIM 进行生产环境推理，或者通过领先的 AI 云服务和推理合作伙伴部署模型，包括 Baseten（Embed、Rerank、Parse）、DeepInfra（Embed 1B、Embed 8B、Rerank 8B）和 FriendliAI（Embed 1B、Embed 8B、Rerank 8B）。Developers can access the open weights, data, training recipes, and more on Hugging Face, deploy the NVIDIA NIM for production inference, or deploy the models through leading AI Cloud and Inference Partners including Baseten (Embed, Rerank, Parse), DeepInfra (Embed 1B, Embed 8B, Rerank 8B), and FriendliAI (Embed 1B, Embed 8B, Rerank 8B).

• Nemotron-3.5-Parse: Hugging Face, NIM
• Nemotron-3.5-Embed-8B-BF16: Hugging Face, NIM
• Nemotron-3.5-Embed-1B-BF16: Hugging Face, NIM
• Nemotron-3.5-Embed-1B-NVFP4: Hugging Face, NIM
• Nemotron-3.5-Rerank-8B-BF16: Hugging Face, NIM

对于使用 NVIDIA AutoModel 进行领域适配，我们提供多模态嵌入和重排序微调支持，帮助团队根据自身的文档集合和相关性要求调整模型。请从 Nemotron 嵌入和重排序微调方案开始使用。For domain adaptation with NVIDIA AutoModel, we provide fine-tuning support for multimodal embedding and reranking, enabling teams to adapt the models to their own document collections and relevance requirements. Get started with the Nemotron embedding and reranking fine-tuning recipes.

根据工作流选择模型：使用 Parse 从文档中提取结构化信息，使用 Embed 搜索文本和图像，使用 Rerank 优化检索结果的排序。将它们组合起来，构建能够为 AI 智能体提供更有力证据的检索流程。Choose a model based on your workflow: Parse for extracting structured information from documents, Embed for searching text and images, and Rerank for improving the ordering of retrieved results. Combine them to build retrieval pipelines that deliver stronger evidence to AI agents.

致谢及主要贡献者 Acknowledgments and main contributors

Nemotron 3.5 Retriever 的发布是 NVIDIA 内多个团队和众多人员通力合作的成果，其中包括主要贡献者。Nemotron 3.5 Retriever release is the result of amazing collaboration between many teams and people across NVIDIA, including main conributors

罗奈·阿克、尤亨·巴巴欣、贝赫扎德·博鲁杰尔迪安、蔡佳瑞、瑞安·切斯勒、佐伊·福尔（cw）、詹妮弗·哈伊杜克、史蒂夫·韩、奥利弗·霍尔沃西、伊莎贝尔·赫尔斯曼、刘博、加布里埃尔·莫雷拉、埃文·奥尔德里奇、拉德克·奥斯穆尔斯基、丹尼斯·奥夫钦尼科夫、胡利奥·佩雷斯、康·范、亚历克斯·理查兹、詹姆斯·荣、柳智贤、索海尔·萨希、雷吉斯·夏维、肖恩·索达、马雷克·瓦夫若斯、雅库布·扎克热夫斯基、露西·张、张宇禾。Ronay Ak, Yauhen Babakhin, Behzad Boroujerdian, Jiarui Cai, Ryan Chesler, Zoe Faul (cw), Jennifer Hajduk, Steve Han, Oliver Holworthy, Isabel Hulseman, Bo Liu, Gabriel Moreira, Even Oldridge, Radek Osmulski, Denis Ovchinnikov, Julio Perez, Khang Pham, Alex Richards, James Rong, Jihyeon Ryu, Sohail Sahi, Regis Schiavi, Sean Sodha, Marek Wawrzos, Jakub Zakrzewski, Lucy Zhang, Yuhe Zhang.