NVIDIA Open-Sources Nemotron 3.5 Super VL: 120B Multimodal Agent Model

10 月 15 日，NVIDIA 发布并开源多模态模型 Nemotron 3.5 Super VL，主打将视觉与文本理解合并进单一模型以支撑长时间运行的 Agent 任务，权重、数据集与训练配方全部公开。

已确认

• 架构与规格：总参数 120B、每 token 激活 12B 的 MoE，外加视觉编码器；支持文本、图像、文档与视频理解，以及推理和工具调用，上下文达 1M token（Artificial Anlys 数据）
• Hugging Face 上放出 base、BF16、NVFP4 量化版本及 LiLiCorr 投机解码配置；据 NVIDIA 团队成员 gordicaleksa，开放内容包括 Base BF16 与后训练 BF16/NVFP4 权重、数据集和训练配方（kuchaev 帖）
• 性能：NVIDIA 引用 Artificial Analysis 数据称输出速度约 640 token/s，在 100K 输入 token 下比同智能指数区间开源模型快约 5 倍；在 4 块 B200 上完成 10,000 个 SWE-Bench Pro 编码任务约需 12 小时、速度最高提升 4 倍；gordicaleksa 给出 SWE-bench Pro 51.78%、吞吐 2.34 倍的数据
• 榜单：Artificial Analysis 称其 Intelligence Index 得 26 分，登顶美国开源权重模型；kuchaev 称多数基准超越自家 Ultra-3.0
• 生态：Baseten（Model APIs 与 Loops）与 Together AI 提供 day 0 支持，Together 强调 NVFP4 KV 缓存；官方演示包括识别店铺安全隐患、分析图表、查账的视觉调查 Agent，可在单台 DGX Spark 上运行
• NVIDIA 同期还发布了 74B-A9B 的 Nemotron 3.5 Bolt（2-bit 量化塞进 24GB 显存跑本地 Agent），以及 Lightning 模型（gordicaleksa 提及）
• NVIDIA 还展示了文档智能、网络安全、金融、科学实验、旅行五大行业后训练后准确率均有提升

为什么重要

• 单模型统一文本与视觉理解，减少了传统方案在独立文本模型与视觉模型之间传递上下文的开销，直接面向多模态 Agent 工作流这一高热方向
• 全量开源（权重+数据+配方）加上高吞吐与单卡可运行特性，降低本地与云端部署门槛，对开源多模态模型格局形成冲击

2026-10-15 ~ 2026-10-16 · 28 related posts

Full story(3 episodes)→

• Episode 1: NVIDIA Launches Nemotron 3.5: Super 120B and Lightning 30B Dual Models(2026-10-15, 5 posts)
• Episode 2: NVIDIA Open-Sources Nemotron 3.5 Super VL: 120B Multimodal Agent Model(2026-10-15, 28 posts)
• Episode 3: NVIDIA Nemotron 3.5 Super VL tops US open-weight models(2026-10-15, 2 posts)

• NVIDIA releases Nemotron 3.5 Super VL 120B-A12B open model family — jinnyjuice · 2026-10-15
• NVIDIA Nemotron 3.5 Super VL, an open-weight visual reasoning model, is live on Together AI — togethercompute · 2026-10-15
• NVIDIA open-sources Nemotron 3.5 Bolt 74B VLM that fits local agents in 24GB via native 2-bit quant — jacek2023 · 2026-10-15
• NVIDIA explains how Nemotron 3.5 Super VL unifies text and vision in one model — NVIDIAAI · 2026-10-15
• NVIDIA: Nemotron 3.5 Super VL completes 10,000 coding tasks up to 4x faster — NVIDIAAI · 2026-10-15
• NVIDIA shows post-training accuracy gains for Nemotron 3.5 Super VL across five domains — NVIDIAAI · 2026-10-15
• [source] NVIDIA open-sources Nemotron 3.5 Super VL: 120B MoE with 5x faster agentic throughput — NVIDIAAI · 2026-10-15
• NVIDIA charts Nemotron 3.5 Super VL at 5x faster throughput among open-weights models — NVIDIAAI · 2026-10-15
• [source] NVIDIA releases open Nemotron 3.5 Super VL: 120B total params, 12B active for vision agents — nvidia · 2026-10-15
• NVIDIA ships Nemotron-3.5-Super-VL: 120B/12B-active VLM beats Ultra-3.0 on benchmarks — kuchaev · 2026-10-15
• NVIDIA open-sources Nemotron 3.5 Super VL, a 120B multimodal model built for long-running agents — NVIDIA Developer · 2026-10-15
• [source] NVIDIA open-sources Nemotron-3.5-Super-VL: 120B weights, datasets, recipes and tech report — kuchaev · 2026-10-15
• NVIDIA's open 120B Super VL model powers a visual agent that audits your AI bills — NVIDIA Developer · 2026-10-15
• NVIDIA's open Nemotron 3.5 Super VL (120B-A12B) lands on Baseten day 0 — baseten · 2026-10-15
• Nemotron 3.5 Super VL Goes Live on Baseten with LoRA Fine-tuning SDK — baseten · 2026-10-15
• NVIDIA ships Nemotron 3.5 Super & Lightning: 120B MoE VLM scores 51.78% on SWE-bench Pro — gordic_aleksa · 2026-10-15
• NVIDIA ships two open Nemotron 3.5 Bolt models: a 74B-A9B MoE and a mobile-friendly E4B — HankYeomans · 2026-10-15
• NVIDIA's Nemotron 3.5 Super VL scores 26 to become top US open-weight model; Bolt 74B nears Ultra at 1/7 the params — ArtificialAnlys · 2026-10-15
• NVIDIA ships two Nemotron-3.5 open models in one day: Super for H100, Bolt for RTX — kuchaev · 2026-10-16
• NVIDIA squeezes a 74B open model into 24GB VRAM, plus a 120B agent model that reads PDFs — mawcopolow · 2026-10-16

8 near-duplicate retellings: togethercompute · togethercompute · kuchaev · testingcatalog · baseten · NVIDIAAI · IanAndrewsDC · yoavartzi