NVIDIANVIDIA · 当前代current generation Nemotron 3.5Nemotron 3.5(2026-11)
NemotronNemotron模型家族谱系卡model family lineage card
NVIDIA 的开放权重 agent 家族:三档 MoE 覆盖单机到集群,主打长上下文、推理与本地 agent。它的价值不在“又一个开源模型”,而在把 NVIDIA 的推理栈、NIM 与硬件适配绑在一起;如果你不用 NVIDIA 硬件,收益会打折。NVIDIA’s open-weight agent family: three MoE tiers from a single machine to a cluster, pitched at long context, reasoning and local agents. Its value is not “another open model” but the coupling with NVIDIA’s inference stack, NIM and hardware; without NVIDIA hardware the benefit is smaller.
本站尚未收录 Nemotron 的量化对比:三档尺寸跨度大,GGUF 生态与下载量还在早期,未达到本站约 30 万下载的门槛。No Nemotron quantisation comparison yet: the three tiers span very different sizes and the GGUF ecosystem is still early, below this site’s ~300k-download threshold.
中文生态China availability 国产 / 国内可直接用 / 备案状态 · 快照domestic / direct access / CAC filing · snapshot 2026-11-21
- 出品属地Origin 国外出品foreign-made 出品方为 NVIDIA(美国)。Made by NVIDIA (United States).
- 国内下载Direct access 直连未资料direct unknown 本次未做中国大陆网络探测。No mainland-China network probe was run.
- 备案状态CAC filing 备案未资料filing unknown:未检索备案名单。Filing registry not searched.
备案信息按网信办《生成式人工智能服务已备案信息》公示名单抄编号;「国内可直连」是在本地网络实际打开过一次的记录,不是厂商承诺。Filing status is quoted from the CAC public registry by number. "Direct access" means the page opened once from a mainland network — not a vendor promise.
- NVIDIA Nemotron 官方页:NVIDIA Nemotron official page: https://www.nvidia.com/en-us/ai/nemotron/
代次Generations
-
Nemotron 3.5 Lightning 30B-A3BNemotron 3.5 Lightning 30B-A3B2026-11
轻量档:30B 总参 / 3B 激活 MoE,面向单机与低延迟 agent;官方给出 1M 上下文。Lightweight tier: 30B total / 3B active MoE, aimed at single-machine and low-latency agents; the official page states a 1M context.
官方页面 ↗official page ↗ -
Nemotron 3.5 Super 120B-A12BNemotron 3.5 Super 120B-A12B2026-11
中档:120B 总参 / 12B 激活,适合多卡推理与较复杂 agent 工作流。Mid tier: 120B total / 12B active, suited to multi-GPU inference and more complex agent workflows.
官方页面 ↗official page ↗ -
Nemotron 3.5 Ultra 550B-A55BNemotron 3.5 Ultra 550B-A55B2026-11
旗舰:550B 总参 / 55B 激活,面向大集群与最高质量档;本地部署门槛很高。Flagship: 550B total / 55B active, aimed at large clusters and the highest-quality tier; the local-deployment bar is very high.
官方页面 ↗official page ↗
尺寸档Size tiers
- 30B-A3B(单机/低延迟)30B-A3B (single machine / low latency)
- 120B-A12B(多卡)120B-A12B (multi-GPU)
- 550B-A55B(集群)550B-A55B (cluster)
许可Licences
- nvidia-open-model-licenseNVIDIA Nemotron Open Model License / OpenMDW-1.1(以模型卡为准)NVIDIA Nemotron Open Model License / OpenMDW-1.1 (per model card) 许可原文 ↗licence text ↗
官方能力声明(带出处)Official capability claims (with sources)
- NVIDIA 官方 Nemotron 页面列出 Lightning 30B-A3B、Super 120B-A12B、Ultra 550B-A55B 三档,并标注 1M 上下文与开放模型许可。NVIDIA’s official Nemotron page lists Lightning 30B-A3B, Super 120B-A12B and Ultra 550B-A55B, with a 1M context and an open-model licence.出处 ↗source ↗
以上都来自厂商官方页面或官方自评基准的原文;厂商自己的横评,这里照原文转述,不当结论。The above are quotations of official pages or transcriptions of vendor-run benchmarks; this site does not endorse cross-model conclusions.
出处Sources
- NVIDIA · Nemotron LLM 官方信息页:NVIDIA · Nemotron LLM official information page: https://www.nvidia.com/en-us/ai-data-science/foundation-models/nemotron/llm-info/
- NVIDIA · Nemotron 模型卡与许可入口:NVIDIA · Nemotron model cards and licence entry: https://huggingface.co/nvidia
资料时间:2026-11-21 · 没实测的字段就留空,不猜。Data as of 2026-11-21 — fields we did not measure are left blank.