NVIDIA Nemotron 3.5 Super VL 通过多模态推理为专业智能体赋能

常驻运行的代理正承担越来越长、越来越复杂的企业工作流。单个请求可能涉及阅读文档、跨多个来源核查信息、遵循公司政策，以及调用多个工具。Always-on agents are taking on longer, more complex enterprise workflows. A single request can involve reading a document, checking information across sources, following company policy, and calling several tools.

NVIDIA Nemotron 3.5 Super VL 是一款开放、可定制的多模态模型，适用于这些专业工作流。其 120B 参数的混合专家（MoE）架构会为每个 token 激活 12B 参数，兼顾模型容量与推理效率。NVIDIA Nemotron 3.5 Super VL is an open, customizable multimodal model for these specialized workflows. Its 120B-parameter mixture-of-experts (MoE) architecture activates 12B parameters per token, balancing model capacity with inference efficiency.

在单张 GPU 上，与同尺寸类别的其他开放模型相比，该模型在智能体任务上的吞吐量最高可提升 5 倍。On a single GPU, this model delivers up to 5x higher throughput for agentic tasks compared to other open models in its size category.

该模型将推理、指令遵循和工具使用能力与文本、图像、文档及视频理解能力相结合。通过将这些能力整合到一个模型中，团队无需分别部署文本模型和视觉语言模型，也无需在模型之间传递任务上下文，即可处理文本和视觉任务。The model combines reasoning, instruction following, and tool use with text, image, document, and video understanding. By combining these capabilities into one model, teams can handle text and visual tasks without deploying separate text and vision-language models or passing task context between them.

开发者可以使用自己的数据、工具、策略和工作流，通过 NVIDIA NeMo 对 Nemotron 3.5 Super VL 进行后训练。本文介绍如何基于这一基础构建专用企业智能体，并阐述其背后的训练与推理技术。Developers can post-train Nemotron 3.5 Super VL with NVIDIA NeMo using their own data, tools, policies, and workflows. This post explains how to use that foundation to build specialized enterprise agents and describes the training and inference techniques that power it.

智能体需要理解自身的任务及其运行环境。客服智能体必须遵循公司的退货政策。研究智能体会关联不同来源的证据，而文档智能体则会解读表格、图表和页面布局。Agents need to understand both their tasks and operating environments. A customer-service agent must follow company return policies. Research agents connect evidence across sources, while document agents interpret tables, charts, and page layouts.

Nemotron 3.5 Super VL 将推理能力与视觉理解相结合，适用于多种智能体工作流，包括：Nemotron 3.5 Super VL combines reasoning and visual understanding across a range of agent workflows, including:

• 编码：测试和验证代码，分析错误，并建议修复方案。
  Coding: Test and validate code, analyze errors, and recommend fixes.
• 网络研究：跨网站搜索、整合信息并核实答案。
  Web research: Search across websites, connect information, and verify answers.
• 文档智能：从表格、图表和页面布局中提取信息，以生成摘要和报告。
  Document intelligence: Extract information from tables, charts, and page layouts to produce summaries and reports.
• 网络安全：识别安全漏洞并对威胁进行分级，以确定修复优先级。
  Cybersecurity: Identify security vulnerabilities and triage threats to prioritize remediation.
• 视频理解：提取信息，用于搜索、内容审核和事件分析。
  Video understanding: Extract information for search, content review, and incident analysis.

在模型体系中，Nemotron 3.5 Super VL 负责处理需要推理、使用工具以及同时理解文本和视觉内容的专业多步骤任务。较小的模型（例如 NVIDIA Nemotron 3.5 Lightning）负责处理大批量文本任务，而 NVIDIA Nemotron 3 Ultra 等前沿模型则负责应对复杂的规划与编排工作。Within a system of models, Nemotron 3.5 Super VL handles specialized, multistep work that needs reasoning, tool use, and text and visual understanding together. Smaller models, such as NVIDIA Nemotron 3.5 Lightning, handle high-volume text tasks, while frontier models such as NVIDIA Nemotron 3 Ultra tackle complex planning and orchestration.

针对领域数据、工具和策略定制 Nemotron 3.5 Super VLCustomize Nemotron 3.5 Super VL for domain data, tools, and policies

专业化智能体必须了解组织的运作方式。领域术语、工具接口、操作流程和政策要求都会影响智能体能否成功完成任务。Specialized agents must learn how an organization works. Domain terminology, tool interfaces, operating procedures, and policy requirements shape how agents complete tasks successfully.

来自网络安全、金融、医疗保健等行业的组织正在对 Nemotron 3.5 Super VL 进行后训练，使其在生产工作流中的准确性和效率方面超越前沿模型。例如：Organizations in industries ranging from cybersecurity and finance to healthcare are post-training Nemotron 3.5 Super VL to outperform frontier models in accuracy and efficiency across their production workflows. For example:

• 人工智能数据与技术服务商 Scale AI 为媒体娱乐公司 Skydance Corporation 后训练了 Nemotron 3.5 Super VL，用于自动化供应商信用处理。该工作负载的推理成本将降至原来的 1/40，延迟将降至原来的 1/5。
  Scale AI, an AI data and technology provider, post-trained Nemotron 3.5 Super VL for automated vendor credit processing at Skydance Corporation, a media and entertainment company. Inference cost for this workload will decrease by 40x and latency by 5x.
• Ramp 是一家财务运营平台提供商，正在将收据图像转换为包含商户信息、日期和付款数据的结构化 JSON。Ramp 预计吞吐量最高可提升 50%，延迟可降至原来的十分之一。
  Ramp, a financial operations platform provider, is converting receipt images into structured JSON containing merchant information, dates, and payment data. Ramp expects up to 50% higher throughput and 10x lower latency.
• AI21 和 Alice 对 Nemotron 3.5 Super VL 进行了后训练，使其适用于威胁情报和漏洞分析，准确率提升了 41%。
  AI21 and Alice post-trained Nemotron 3.5 Super VL for threat intelligence and vulnerability analysis, improving accuracy by 41%.

团队可以使用 NVIDIA NeMo 开源库对模型进行后训练，以满足这些要求。开发者可以使用 NVIDIA NeMo AutoModel 进行监督微调（SFT）和低秩适配（LoRA），使模型具备特定领域的行为能力。强化学习（RL）则能提升模型在反映组织工作流的智能体环境中的表现。Teams can use NVIDIA NeMo open libraries to post-train models to meet these requirements. Developers can use NVIDIA NeMo AutoModel for supervised fine-tuning (SFT) and low-rank adaptation (LoRA) to teach domain-specific behavior. Reinforcement learning (RL) improves performance in agentic environments that reflect organizational workflows.

除了 Nemotron 3.5 Super VL，NVIDIA 还将在 Hugging Face 上发布开放数据集。这些数据集包括智能体工具使用和对话数据、金融推理与工具调用数据集、指令遵循和聊天数据、长上下文转码与翻译数据、公平性数据，以及 40 个视觉语言数据集。这些数据集可用于补充专有数据，以支持 SFT 和 RL。Alongside Nemotron 3.5 Super VL, NVIDIA is releasing open datasets on Hugging Face. These include agentic tool-use and conversational data, finance reasoning and tool-calling datasets, instruction-following and chat data, long-context transduction and translation data, fairness data, and 40 vision-language datasets. These datasets can supplement proprietary data for SFT and RL.

对于多模态数据，可先从 NeMo Data Designer 中有关图表问答和长文档理解的资源入手。探索 NVIDIA NeMo RL 和 NVIDIA NeMo Gym，了解强化学习和基于环境的评估。For multimodal data, start with NeMo Data Designer resources for chart question answering and long-document understanding. Explore NVIDIA NeMo RL and NVIDIA NeMo Gym for RL and environment-based evaluation.

组织可以在受支持的本地、数据中心或云环境中调整并部署该模型的开放权重，以满足自身的基础设施和数据要求。有关定制资源，请参阅 GitHub 上的 Nemotron 3.5 Super VL 使用指南。Organizations can adapt and deploy the model’s open weights in supported local, data center, or cloud environments to meet their infrastructure and data requirements. For customization resources, see the Nemotron 3.5 Super VL usage cookbook on GitHub.

在一个工作流中结合文本理解与视觉理解Combine text and visual understanding in one workflow

Nemotron 3.5 Super VL 将文本、图像、文档和视频理解与推理及指令遵循能力相结合，从而回答有关这些输入的问题。Nemotron 3.5 Super VL combines text, image, document, and video understanding with reasoning and instruction following to answer questions across these inputs.

其视觉训练涵盖图表、表格、光学字符识别（OCR）、多页文档和视频理解。两阶段 SFT 流程用于训练语言和视觉组件，并结合视觉证据、推理、对话、安全、长上下文数据和智能体轨迹。Its visual training covers charts, tables, optical character recognition (OCR), multipage documents, and video understanding. A two-stage SFT pipeline trains the language and vision components, combining visual evidence with reasoning, chat, safety, long-context data, and agent trajectories.

在文档智能领域，同一个模型可以解读图表，将其与文档其他部分的表格关联起来，并对两者进行总结。开发者可以根据自己的文档和评估标准定制工作流程。For document intelligence, the same model can interpret a chart, connect it with a table elsewhere in a document, and summarize both. Developers can tailor the workflow to their own documents and evaluation criteria.

使用 NeMo Data Designer 生成的多样化训练数据，帮助模型更好地理解复杂图表。在图表推理基准测试 CharXiv 上，该模型得分为 80.8%，而 Nemotron 3 Nano Omni 的得分为 63%。结合 Python 代码解读后，得分达到 84%。有关方法和经验总结，请阅读图表理解深度解析以及开源图表问答流程。Diverse training data generated with NeMo Data Designer helped the model better understand complex charts. On CharXiv, a chart-reasoning benchmark, the model scored 80.8%, compared with 63% for Nemotron 3 Nano Omni. With Python code interpretation, it reached 84%. For the approach and lessons learned, read the chart-understanding deep dive and the open-source chart question-answering pipeline.

该模型的构建得到了 Nemotron Coalition 组织的支持和专业知识。例如，Perplexity 提供的高质量数据集帮助改进并精简了模型的中期训练。Prime Intellect 提供的环境则助力构建更优秀的教师模型。This model was built with support and expertise from Nemotron Coalition organizations. For instance, high-quality datasets from Perplexity helped improve and streamline the model’s mid-training. Environments from Prime Intellect enabled building better teacher models.

Nemotron 3.5 Super VL 也是 OCRBench v2 排行榜上的顶尖开源模型之一。Nemotron 3.5 Super VL is also one of the top open models on the OCRBench v2 leaderboard.

代理任务中的准确性与效率Accuracy and efficiency across agent tasks

Nemotron 3.5 Super VL 的评测涵盖 Hermes、LangChain、OpenClaw、OpenCode、Pi、Codex 和 Claude。除了编程任务外，智能体运行框架提供商还会使用自己的评测来测试各种各样的智能体任务。在 Hermes Bench 和 LangChain 行为评测中，Nemotron 3.5 Super VL 的得分高于 Nemotron 3 Ultra。Evaluations of Nemotron 3.5 Super VL span Hermes, LangChain, OpenClaw, OpenCode, Pi, Codex, and Claude. In addition to coding tasks, agent harness providers use their own evaluations to test a wide range of agent tasks. On Hermes Bench and LangChain behavioral evaluations, Nemotron 3.5 Super VL scores higher than Nemotron 3 Ultra.

对于持续运行的智能体，效率决定了部署能够完成多少工作。吞吐量、令牌用量、工具调用次数，以及得出正确结果所需的步骤，都会影响任务完成时间。For always-on agents, efficiency determines how much work a deployment can complete. Throughput, token usage, tool calls, and the steps needed to reach a correct result all affect task completion time.

该模型还经过优化，可在 NVIDIA DGX Spark 上本地运行。EXO Labs 的独立评估将 Nemotron 3.5 Super VL 列入准确率与任务完成时间的帕累托前沿。This model was also optimized to run locally on NVIDIA DGX Spark. An independent evaluation by EXO Labs placed Nemotron 3.5 Super VL on the Pareto frontier for accuracy and task completion time.

Nemotron 3.5 Super VL under the hood

该模型建立在坚实的基础之上，并采用混合架构、推测解码、上下文复用和混合精度，以减少多次智能体调用中的推理工作量。The model builds on a strong foundation and uses a hybrid architecture, speculative decoding, context reuse, and mixed precision to reduce inference work across repeated agent calls.

构建一个强大的基础模型Building a strong base model

Nemotron 3.5 Super VL 是从 Nemotron 3 Ultra 蒸馏而来的。Ultra 的编码能力与蒸馏过程中使用的 SkewReverse 损失相结合，有助于提高编码准确率。Nemotron 3.5 Super VL was distilled from Nemotron 3 Ultra. Ultra’s coding capabilities, combined with the SkewReverse loss used during distillation, helped improve coding accuracy.

在后训练阶段，多教师在线策略蒸馏（MOPD）融合了文本和视觉领域专业教师的指导。模型生成回答后，教师会对其视频推理与跟踪、视觉工具使用以及对齐能力进行评分。During post-training, multi-teacher on-policy distillation (MOPD) combines guidance from specialized teachers across text and vision. The model generates responses, and the teachers score them for video reasoning and tracking, vision tool use, and alignment.

通过增加关于智能体操作、代码仓库和软件工程任务的示例训练，模型在智能体任务上的表现有所提升。在多种评测设置下，Nemotron 3.5 Super VL 在 SWE-bench 和 Terminal-Bench 上的表现均优于 Nemotron 3 Super。Additional training on examples of agent actions, code repositories, and software engineering tasks improved the model’s performance on agent tasks. Nemotron 3.5 Super VL outperforms Nemotron 3 Super on SWE-bench and Terminal-Bench across multiple evaluation setups.

使用高质量的自然数据和合成数据进行持续预训练，提升了模型对印度语言的能力，并使后训练后的总体得分更高。此阶段新增了 100 亿个印度语言 token，数据取自 Nemotron-pretraining-Indic-v1 的 12 种语言子集以及自然印度语文本。Continued pretraining with high-quality natural and synthetic data improved Indic-language capabilities and contributed to higher overall scores after post-training. This stage added 10B Indic tokens, drawn from a 12-language subset of Nemotron-pretraining-Indic-v1 and natural Indic text.

NVIDIA 正在发布完整数据集，该数据集包含约 2.8 万亿个 token，涵盖 22 种印度语言，由 NVIDIA 与印度理工学院马德拉斯分校及其 AI4Bharat 实验室共同开发。以下结果基于经过后训练的 Nemotron 3.5 Super VL。NVIDIA is releasing the full dataset, which contains about 2.8T tokens across 22 Indian languages and was developed with IIT Madras and its AI4Bharat lab. The following results are based on the post-trained Nemotron 3.5 Super VL.

开发者可以基于这一基础，使用 NeMo 构建并定制 Nemotron 3.5 Super VL，以适配其应用，并根据自身的数据、工具和工作流程调整模型的推理能力和智能体能力。Developers can use NeMo to build on this foundation and customize Nemotron 3.5 Super VL for their applications, adapting its reasoning and agentic capabilities to their data, tools, and workflows.

LiLiCorr：一种开放式推测解码技术LiLiCorr: An open speculative decoding technique

LiLiCorr 通过在整个草稿范围内联合评估候选词元，改进了并行推测解码，捕捉到独立词元预测所遗漏的依赖关系。一个轻量级网络只需单次前向计算即可得出兼容性分数，从而生成更连贯的草稿，最后仅需执行低成本的顺序选择步骤。LiLiCorr improves parallel speculative decoding by jointly evaluating token candidates across an entire draft, capturing dependencies that independent token predictions miss. A lightweight network computes compatibility scores in a single pass to produce more coherent drafts, leaving only a low-cost, sequential final selection step.

在 DGX Spark 上，LiLiCorr 是使用 Nemotron 3.5 Super VL 进行低延迟推理的推荐草稿器。与不使用推测解码相比，它的 token 生成速度最高可提升 2.3 倍。On DGX Spark, LiLiCorr is the recommended drafter for low-latency inference with Nemotron 3.5 Super VL. It delivers up to 2.3x faster token generation than decoding without speculation.

NVFP4 与混合精度NVFP4 and mixed precision

该模型使用 NVIDIA Model Optimizer 进行训练后量化和量化感知蒸馏（QAD）。NVFP4 配方在 NVIDIA Blackwell 架构上使用 4 位权重和激活值。同一检查点在非 Blackwell 平台上运行时使用 4 位权重和 16 位激活值。敏感组件的精度需单独选择。The model uses NVIDIA Model Optimizer for post-training quantization and quantization-aware distillation (QAD). The NVFP4 recipe uses 4-bit weights and activations on the NVIDIA Blackwell architecture. The same checkpoint runs on non-Blackwell platforms with 4-bit weights and 16-bit activations. Precision is selected separately for sensitive components.

NVFP4 可以缩小模型体积，而 QAD 则帮助模型在较低精度下保持准确率。在 QAD 过程中，量化后的模型会学习匹配原始 BF16 模型的输出，从而弥补量化造成的精度差距。QAD 还支持将键值（KV）缓存存储为 NVFP4，而非 FP8，同时仍能达到准确率目标。由于 QAD 在训练期间完成，因此不会增加推理开销。NVFP4 makes the model smaller, and QAD helps it retain its accuracy at lower precision. During QAD, the quantized model learns to match the outputs of the original BF16 model, which closes the accuracy gap that quantization introduces. QAD also supports storing the key-value (KV) cache in NVFP4 instead of FP8 while still meeting accuracy targets. Because QAD happens during training, it adds no inference overhead.

多词元预测Multi-token prediction

多 token 预测（MTP）为 Nemotron 3.5 Super VL 配备了一个内部草拟器，可在一次前向传播中为主模型提出多个 token 供其验证。一次接受多个 token，可以减少生成响应所需的高计算量验证步骤。Multi-token prediction (MTP) gives Nemotron 3.5 Super VL an internal drafter that proposes several tokens for the main model to verify in one forward pass. Accepting multiple tokens at once reduces the number of compute-intensive verification steps needed to generate a response.

专门用于提升 MTP 的训练阶段进一步使起草模型与最终模型保持一致。A dedicated MTP-boosting training phase further aligned the drafter with the final model.

跨智能体轮次的上下文复用Context reuse across agent turns

智能体会反复将系统指令、工具定义和对话历史传递给模型。前缀缓存会复用这些共享输入的计算结果。Agents repeatedly pass system instructions, tool definitions, and conversation history back to the model. Prefix caching reuses computation from those shared inputs.

该模型的服务优化同时保留注意力缓存和 Mamba 循环状态。增加 Mamba 检查点可减少为推测解码恢复输入所需的回放操作，而内部检查点机制则可避免在处理提示词时额外运行模型。The model’s serving optimizations preserve both attention caches and Mamba recurrent states. Additional Mamba checkpoints reduce the replay needed to recover inputs for speculative decoding, while internal checkpointing avoids extra model passes during prompt processing.

借助 NVIDIA NIM 和 NVIDIA Dynamo 简化生产环境部署Simplify production deployment with NVIDIA NIM and NVIDIA Dynamo

NVIDIA NIM 将模型与经过优化的推理引擎、经过验证的配置和标准 API 打包，适用于云端和数据中心部署。NIM 持续获得安全性、功能和性能更新。可通过 NVIDIA AI Enterprise 获取企业级支持。立即开始使用 NIM。NVIDIA NIM packages models with optimized inference engines, validated configurations, and standard APIs for cloud and data center deployment. NIM receives ongoing security, feature, and performance updates. Enterprise support is available through NVIDIA AI Enterprise. Get started with NIM.

NVIDIA Dynamo 支持分布式推理优化，以提升吞吐量并降低延迟，包括感知 KV 缓存的路由、KV 缓存管理和解耦式服务。NVIDIA Dynamo supports distributed inference optimizations to increase throughput and reduce latency, including KV-cache-aware routing, KV cache management, and disaggregated serving.

在 16 块 NVIDIA B200 GPU 上，以每位用户每秒 50 个 token 的速度运行时，采用 KV 感知路由、会话亲和性和 KV 缓存卸载，最多可服务 8 倍的用户，同时将首 token 延迟降低 85%。立即开始使用 Dynamo 部署方案。On 16 NVIDIA B200 GPUs at 50 tokens per second per user, KV-aware routing with session affinity and KV cache offload serves up to 8x as many users while reducing first-token latency by 85%. Get started with the Dynamo deployment recipes.

Get started with Nemotron 3.5 Super VL

构建具备多模态推理和工具使用能力的专用智能体。针对专业工作流定制模型，然后在受支持的环境中部署。Build specialized agents with multimodal reasoning and tool use. Customize the model for specialized workflows, then deploy it in a supported environment.

该版本包含基础、后训练及 NVFP4 量化检查点，以及训练配方、数据和强化学习环境。The release includes base, post-trained, and NVFP4-quantized checkpoints with training recipes, data, and reinforcement-learning environments.

模型详情请参阅 Nemotron 3.5 技术报告；部署说明请参阅 Hugging Face 或 ModelScope。有关托管模型的访问方式，请参阅 OpenRouter。See the Nemotron 3.5 technical report for model details and Hugging Face or ModelScope for deployment instructions. For hosted model access, see OpenRouter.

合作伙伴生态系统Partner ecosystem

合作伙伴支持针对特定领域的工作流定制和部署 Nemotron 3.5 Super VL。Partners support customization and deployment of Nemotron 3.5 Super VL for domain-specific workflows.