英伟达发布全新开源模型：赋能企业与 PC 本地端运行 AI Agent

英伟达于周四发布了专为全天候运行的 Agent 和边缘计算设计的新型开源权重模型。

其中，Nemotron 3.5 Super VL 是一款拥有 1200 亿参数（120B）的大型模型，专为企业级 Agent 场景打造。尽管其参数规模小于 Nemotron 3 Ultra 模型，但英伟达表示，它在指令遵循和工具调用任务中能够提供相当的准确率。

而在另一端，英伟达还推出了两款轻量级的 Bolt 模型。这些模型旨在新一代 AI PC（搭载英伟达 RTX Spark 芯片）以及搭载 DGX Spark 的工作站上本地运行。面向消费者、程序员和企业知识工作者的此类 PC，目前正在通过微软和戴尔等合作伙伴推向市场。

“任何希望在英伟达基础设施上运行 Agent 的组织，现在都能根据自身层级获得开源且经过优化的模型，而无需在能力与可部署性之间做出妥协，”英伟达开源模型与软件副总裁 Kari Briski 在发布这些模型的官方博客文章中写道。

英伟达声称，在针对特定 Agent 任务进行后训练时，Super VL 完成任务的速度快于 Anthropic 和 OpenAI 的竞争模型（如 Opus 5、GPT 6.1 Sol、Luna、Astra）。这些任务包括文档智能、财务会计、科学实验、医疗保险理赔以及安全攻击防御。

尽管每当科技巨头发布新模型时都会有此类宣传，但英伟达列出了一批正在使用 Super VL 的重量级企业客户作为背书，其中包括 Skydance、安永（EY）、Ramp、Crowdstrike、Perplexity 以及 Salesforce。

当然，这并不是英伟达首次发布与闭源前沿实验室竞争的开源权重模型——这些实验室既是它的客户，也是它的投资对象。这家 GPU 巨头已经拥有完整的 Nemotron 模型家族，覆盖从多模态人工智能到推理、代码编写、数学计算以及企业自动化等各个领域。

然而，这些快速发展的闭源 AI 实验室目前正开始寻求英伟达之外的其他芯片和 AI 系统供应商（例如 AWS 的 Trainium 和 Cerebras）来满足算力需求，甚至在着手自研芯片。

因此，英伟达正在全力押注企业希望在私有数据中心运行自主开源模型的趋势。“与闭源前沿模型不同，企业组织可以基于自身数据对 Super VL 进行后训练，将其部署在自有基础设施上，并完全拥有自己构建的成果，”Briski 写道。

不过，随着体量更小的本地 PC 模型的发布，英伟达也在试图推动处于萌芽阶段的消费级 AI 助手趋势。尽管初创公司和大型科技公司似乎每天都在推出新的 AI 助手，但对隐私心存顾虑的消费者可能并不愿意让这些 Agent 过深地介入生活，从而获取关于家庭、健康、财务等私密信息。

因此，英伟达希望通过打造能够本地运行模型与 Agent 的设备，开辟销售搭载其芯片的 PC 的新市场，同时 PC 用户“可以委派日常项目……而无需将这些工作负载上传到云端，”Briski 写道。

在两款 Bolt 模型中，Nemotron 3.5 Bolt 74B-A9B 24GB 是一款拥有 740 亿参数的模型，专为本地 Agent 任务设计。它旨在平衡速度与推理能力，以处理多步推理、工具调用和网页浏览。而体量最小的 Bolt E4B 则面向对话式 Agent。

英伟达表示，Super VL 和 Bolt 74B-A9B 目前已正式上线，Bolt E4B 将于本月晚些时候推出。

Nvidia on Thursday launched new open-weight models designed for always-on agents and edge computing.

The Nemotron 3.5 Super VL is a large, 120-billion-parameter model geared for agents used by the enterprise. While it uses fewer parameters than the Nemotron 3 Ultra model, Nvidia says it can deliver similar accuracy in instruction-following and tool-use tasks.

At the other end of the spectrum, the company also launched two lightweight Bolt models. These are intended to run locally on the new AI PCs (running the Nvidia RTX Spark chip) and workstations running the DGX Spark. Such PCs, aimed at consumers, programmers, and enterprise knowledge workers, are being released right now into the market through partners like Microsoft and Dell.

“Any organization that wants to run agents on Nvidia infrastructure now has an open, optimized model for their tier without having to choose between capability and deployability,” Nvidia’s vice president of Open Models and Software, Kari Briski, wrote in the blog post announcing these models.

Nvidia claims that the Super VL completes tasks faster than competing models from Anthropic and OpenAI (Opus 5, GPT 6.1 Sol, Luna, Astra) when post-trained on specific agentic tasks. These tasks include document intelligence, financial accounting, scientific experimentation, medical insurance claims, and security attack resilience.

While such claims are standard whenever a major tech company announces a new model, Nvidia has backed it up with a list of heavy-hitter companies that it says are using Super VL. These include Skydance, EY, Ramp, Crowdstrike, Perplexity, and Salesforce.

Of course, these aren’t the first open-weight models that Nvidia has released that compete with the proprietary frontier labs — companies that are both its customers and its investments. The GPU giant has an entire family of Nemotron models that do everything from multimodal artificial intelligence to reasoning, coding, math, and enterprise automation.

However, those fast-growing proprietary AI labs are now looking to many other chips and AI system makers beyond Nvidia to power their needs, like AWS’s Trainium and Cerebras, as well as building their own home-grown chips.

So Nvidia is leaning hard into the trend of enterprise companies wanting to run their own open models in their own data centers. “Unlike closed frontier models, organizations can post-train Super VL on their own data, deploy it on their own infrastructure and own what they build,” Briski wrote.

But with the announcement or the smaller, local PC models, it is also trying to push forward the nascent trend of consumer AI assistants. While it seems a new AI assistant is introduced every day by startups and Big Tech alike, privacy-weary consumers may not want to invite such agents so deeply into their lives giving them access to info about their families, health, finances, and so on.

So, while Nvidia hopes to create a new market for itself selling PCs powered by its chips by creating machines that run the models and agents locally, PC users “can delegate everyday projects … without sending those workloads to the cloud,” Briski wrote.

Of the two Bolt models, Nemotron 3.5 Bolt 74B-A9B 24GB is a 74-billion parameter model designed for local agentic tasks. It is intended to balance speed with reasoning to handle multistep reasoning, tool calling, and web browsing. The smallest Bolt E4B is aimed at conversational agents.

Super VL and Bolt 74B-A9B are available now, with Bolt E4B coming later this month, Nvidia says.