内容来源：https://aibusiness.com/generative-ai/open-source-on-device-strategy-behind-nvidia-s-new-ai-models

内容总结：

英伟达本周发布三款全新Nemotron 3.5系列开源模型，包括旗舰级Nemotron 3.5 Super VL（1200亿参数）、Nemotron 3.5 Bolt 74B-A9B 24GB以及Nemotron 3.5 Bolt E4B 4GB。三款模型均完全开源，英伟达同步公开了模型权重、数据集和训练方案，并已上线Hugging Face平台。这是英伟达上月同意收购该开源AI平台以来首次发布新模型，也是继今年8月推出Nemotron 3.5 Lightning之后的又一次开源动作。

据报道，旗舰模型Super VL主打快速任务补全能力，可在“常驻在线”智能体场景下近乎即时地完成任务；Bolt 74B-A9B兼顾性能与速度平衡，E4B则面向对话式智能体设计。分析人士指出，近年来开源模型规模显著增大，Meta于2024年推出的4050亿参数Llama模型即为代表，Omdia分析师Alexander Harrowell表示，“我们所认为的小模型已经变得相当大了”，这一变化有利于企业更好地微调和定制模型，同时市场上开源模型增多也带来了更充分的竞争。

报道认为，英伟达此举意在持续支持开源生态，并延续其在美国AI市场主导开源模型发展的路线，以更好地与中国低成本、高性能的开源权重模型竞争。英伟达本周还将GPU编排与AI工作负载平台Run.ai开源。Harrowell指出，英伟达支持开源生态的动机并非完全纯粹，其同时也着眼于围绕自身GPU构建更广泛的生态体系。他提到，谷歌、亚马逊和OpenAI等英伟达客户均在自研AI芯片，行业正呈现“模型、模型软件与定制芯片深度整合”的趋势。

Harrowell还表示，开源权重模型迭代替换速度很快，已有开发者讨论用Nemotron Bolt直接替换阿里巴巴的Qwen 3.6 35B。他认为，英伟达此次聚焦专用智能体对开源社区意义重大，许多开发者正在寻找优质的代码类开源基础模型以支撑定制化智能体，而新模型的架构支持更好的内存扩展，从而获得更大的上下文窗口，这在智能体AI场景中极具价值。此外，业界关注的另一大问题是AI能否从数据中心迁移至本地PC或Nvidia DGX Spark、苹果Mac Mini等共享设备，瓶颈之一在于企业需重新适应软件在本地计算机运行、而非仅存在于云端或作为Web服务的模式。

中文翻译：

由Google Cloud赞助
选择你的首批生成式AI用例
要开始使用生成式AI，首先应聚焦于能够改善人类信息体验的领域。
新的Nemotron 3.5模型展示了这家人工智能巨头如何致力于强化自身GPU基础设施，同时支持开源社区。
在开源与闭源AI之间日益加速的竞赛中，该厂商以更多开放模型推进其开源路线。
英伟达本周早些时候推出了Nemotron 3.5 Super VL，这是一款1200亿参数的模型，据称可实现快速补全，即AI模型几乎瞬间完成任务的能力，并配合“始终在线”的智能体实现高精度。
除了这款顶级模型外，英伟达还发布了Nemotron 3.5 Bolt 74B-A9B 24GB和Nemotron 3.5 Bolt E4B 4GB。64B版本针对平衡与速度进行了优化，而E4B则专为对话式智能体设计。
三款模型完全开源，英伟达发布了其权重、数据集和训练配方。此次发布之前，英伟达于8月发布了另一款完全开放的模型Nemotron 3.5 Lightning。
新模型已在Hugging Face上线。这是英伟达自上月同意收购该开源AI平台以来发布的首批模型。此次发布凸显了自Meta首次发布其Llama开放权重模型系列以来，开放模型范式在过去几年中的演变。
近年来，开放模型的规模显著增大，Meta于2024年推出的4050亿参数Llama模型便是突出例证。
“我们所谓的小模型已经变得相当大了，”Informa TechTarget旗下Omdia的分析师Alexander Harrowell表示。
他补充说，这一变化应该有利于企业，因为他们可以更好地微调和定制模型。此外，企业现在可以获取更多模型，市场上也有更多开放模型，从而促进更好的竞争。
随着最新Nemotron模型的发布，英伟达持续支持开源倡议，并继续在美国AI市场引领开发开放模型的运动，以更好地与低成本、高性能的中国开放权重模型竞争。
最近的一个例子发生在本周早些时候，英伟达将其GPU编排和AI工作负载平台Run.ai开源。
“他们担心开放AI生态系统的必要性，”Harrowell说，并补充称这一动机并不完全纯粹，因为英伟达也有自身的利益考量。
据Harrowell称，虽然英伟达发布开放模型是为了支持开源社区，但这显然不是唯一原因。
“他们这样做是为了支持围绕其GPU的更广泛生态系统，”他说。
与此同时，英伟达的一些客户拥有自己的AI芯片。例如，谷歌提供其TPU，亚马逊提供Trainium AI加速芯片，而OpenAI正在开发自己的AI芯片Jalapeño。
“你可以将这视为一种[全行业]的整合趋势，”Harrowell说，他指的是随着谷歌、亚马逊和OpenAI开发自己的芯片，它们可以将这些芯片与自己的AI模型进行整合和优化。“模型、模型软件和定制芯片的紧密整合。”
他指出，大多数开放权重模型很快就会被替换，因为公民开发者会从较旧的开放模型转向更新的模型。
“我现在就能看到人们在讨论如何直接用Nemotron Bolt替换Qwen 3.6 35B，”Harrowell说，他指的是中国科技巨头阿里巴巴广受欢迎的开源模型。
Harrowell继续表示，英伟达通过发布最新Nemotron模型聚焦专业化智能体，这对开源社区很重要。
“很多开发者正在寻找一个良好的开源代码基础模型来支持定制化智能体，”他说。他补充说，Nemotron 3.5 Super VL和Bolt模型的架构实现了更好的内存扩展，从而带来更大的上下文窗口，这在智能体AI中很有价值。
与此同时，业界的一大问题是AI能否从数据中心迁移到本地PC或Nvidia DGX Spark和苹果Mac Mini等共享设备上。
Harrowell表示，一个瓶颈是企业需要重新习惯打包软件在本地计算机上运行的理念，而不仅仅是在云端或作为Web服务运行。
Esther Shittu是Informa TechTarget的新闻撰稿人和播客主持人，报道AI软件和系统。

英文来源：

Sponsored by Google Cloud
Choosing Your First Generative AI Use Cases
To get started with generative AI, first focus on areas that can improve human experiences with information.
The new Nemotron 3.5 models show how the AI giant is focusing on reinforcing its own GPU infrastructure while supporting the open source community.
Amid the accelerating race between open and closed AI, the vendor advanced its open source line of attack with more open models.
Nvidia earlier this week introduced Nemotron 3.5 Super VL, a 120-billion-parameter model that it said delivers fast-track completion, or the ability for an AI model to complete a task nearly instantly, and accuracy with "always on" agents.
Along with that top-of-the-line model, Nvidia released Nemotron 3.5 Bolt 74B-A9B 24GB and Nemotron 3.5 Bolt E4B 4GB. The 64B version is optimized for balance and speed, while the E4B is designed for conversational agents.
All three models are fully open, with Nvidia releasing their weights, datasets and training recipes. The release comes after Nvidia released another fully open model, Nemotron 3.5 Lightning, in August.
The new models are available on Hugging Face. They are the first models Nvidia has released since it agreed to acquire the open source AI platform last month. The release highlights how the open model paradigm has evolved over the past few years since Meta first released its Llama open-weight model family.
In recent years, open models have grown markedly larger, highlighted by Meta's 405-billion-parameter Llama model, introduced in 2024.
"What we think of as small models has gotten quite bigger,” said Alexander Harrowell, an analyst at Omdia, a division of Informa TechTarget.
He added that this change should benefit enterprises, because they can better fine-tune and customize models. Moreover, enterprises now enjoy access to more models and more open models are on the market, allowing for better competition.
With the release of the latest Nemotron models, Nvidia is sustaining its backing of open source initiatives and continues to lead the campaign in the U.S. AI market to develop open models to better compete with low-cost, high-performance Chinese open-weight models.
A recent example came earlier this week, when Nvidia made its GPU orchestration and AI workload platform Run.ai open source.
“They are concerned about the need for an open AI ecosystem,” Harrowell said, adding that this motivation is not entirely pure, because Nvidia also has self-interest in mind.
While Nvidia is releasing open models to support the open source community, that clearly is not the only reason, according to Harrowell.
“They’re doing this to support the wider ecosystem around their GPUs,” he said.
At the same time, some of Nvidia’s customers boast their own AI chips. For instance, Google offers its TPUs, Amazon offers Trainium AI accelerator chips, and OpenAI is developing its own AI chip, Jalapeño.
“You could see this as an [industry-wide] trend toward integration,” Harrowell said, referring to the idea that as Google, Amazon and OpenAI develop their own chips, they can integrate those chips and optimize them with their AI models. “Close integration of models, model software and custom chips."
He noted that most open-weight models are quickly replaced as citizen developers switch from older open models to newer ones.
“I can see people literally right now talking about how they're replacing Qwen 3.6 35B with Nemotron Bolt straight away,” Harrowell said, referring to Chinese tech giant Alibaba’s popular open model.
Nvidia’s focus on specialized agents with the release of the latest Nemotron models is important for the open source community, Harrowell continued.
"A lot of developers are looking for a good open source base model for code to support customized agents," he said. He added that the architecture of the Nemotron 3.5 Super VL and Bolt models enables better memory scaling, leading to a bigger context window, which is valuable in agentic AI.
Meanwhile, one of the big questions in the industry is whether AI can be moved from data centers onto local PCs or shared devices such as Nvidia DGX Spark and Apple’s Mac Minis.
One bottleneck is enterprises getting used to the idea of packaged software running locally on the computer again, rather than only in the cloud or as a web service, Harrowell said
Esther Shittu is an Informa TechTarget news writer and podcast host covering AI software and systems.