英伟达发布 Nemotron 3.5 开放模型：1200亿参数多模态智能体模型与两款边缘模型上线
英伟达（NVIDIA）10月15日在官方博客发布 Nemotron 3.5 开放模型家族的三款新模型。领衔的 Nemotron 3.5 Super VL 是1200亿参数的多模态模型，每个 token 激活120亿参数，可处理文本、图像、短视频与 PDF 文档，面向需要持续运行的专用 SI 智能体（agent），支持指令跟随、工具调用与多步骤业务流程，上下文长度最高达100万 token，现已在 Hugging Face、OpenRouter、ModelScope 以及 NVIDIA NIM 上线。
同批发布的还有两款面向本地与边缘设备的 Bolt 模型。Nemotron 3.5 Bolt 74B-A9B 24GB 采用原生2比特量化，可在24GB显存内运行，能在 GeForce RTX 电脑与 DGX Spark 上本地执行智能体任务，现已上线；更小的 Bolt E4B 4GB 定于10月23日推出，面向 GeForce RTX、Jetson 等设备上的实体与对话智能体。
英伟达表示，Super VL 以约为其最大推理模型五分之一的参数量，达到了可比的多模态智能体任务表现，吞吐量最高达同级开放模型的5倍，数据引自 Artificial Analysis 智能指数榜单。三款模型均以允许商用的 OpenMDW-1.1 许可证开放，模型权重、数据集与训练配方同步开放。
Nemotron 是英伟达的开放模型家族，同家族还有面向高吞吐场景的 Lightning 与面向高难度推理任务的 Ultra。
NVIDIA Releases Open Nemotron 3.5 Models: 120B Multimodal Agent Model Plus Two Edge Models
NVIDIA on October 15 expanded its Nemotron open model family with three new models in an official blog post. The headliner, Nemotron 3.5 Super VL, is a 120-billion-parameter multimodal model — 12 billion parameters active per token — that reads text, images, short video and PDF documents. Built for always-on specialized SI agents, it supports instruction following, tool use and multistep business workflows with a context length of up to 1 million tokens, and is available now on Hugging Face, OpenRouter, ModelScope and as an NVIDIA NIM.
Two smaller Bolt models target local and edge deployment. Nemotron 3.5 Bolt 74B-A9B 24GB uses native 2-bit quantization to fit in 24GB of memory and run on GeForce RTX PCs and DGX Spark systems; it is available now. The smaller Bolt E4B 4GB, designed for physical and conversational agents on GeForce RTX and Jetson devices, is scheduled for October 23.
NVIDIA said Super VL matches the performance of its largest reasoning model at roughly one-fifth the parameter count and delivers up to five times the throughput of comparable open models, citing the Artificial Analysis Intelligence Index leaderboard. All three models are released under the OpenMDW-1.1 license, which allows commercial use, with open weights, datasets and training recipes.
The Nemotron family also includes Lightning, built for high-throughput workloads, and Ultra, aimed at the hardest reasoning tasks.
发布：GSM·美国（发布链接）
图片来源：图：位于 Santa Clara 的英伟达（NVIDIA）总部（资料图）。图片来源：Wikimedia Commons，CC BY-SA 4.0。Photo: NVIDIA headquarters in Santa Clara (file photo). Source: Wikimedia Commons, CC BY-SA 4.0.
GSM Global · 本篇英文版 English edition of this story
