NVIDIA发布120B多模态模型Nemotron 3.5 Super VL
NVIDIA Releases Nemotron 3.5 Super VL, a 120B Open Multimodal Agent Model
NVIDIA最新旗舰多模态模型发布,架构创新且性能显著超越前代,适合关注Agent落地与多模态能力的开发者参考。
NVIDIA just released Nemotron 3.5 Super VL, an open vision-language model built for always-on enterprise agents. Nemotron 3.5 Super VL is a 120B-parameter mixture-of-experts model from NVIDIA that activates 12B parameters per token. It reads text, images, documents and video, then reasons, calls tools and writes code in one model. For developers, 1 open checkpoint replaces a separate text model along with a separate vision model. NVIDIA also shipped 2 edge models, Nemotron 3.5 Bolt 74B-A9B and Bolt E4B.
NVIDIA 刚刚发布了 Nemotron 3.5 Super VL,这是一个专为始终在线的企业智能体构建的开放视觉-语言模型。Nemotron 3.5 Super VL 是 NVIDIA 推出的一个拥有 120B 参数的混合专家(MoE)模型,每个 token 激活 12B 参数。它能够阅读文本、图像、文档和视频,并在单一模型中进行推理、调用工具和编写代码。对于开发者而言,1 个开放检查点替代了单独的文本模型和单独的视觉模型。NVIDIA 还推出了 2 款边缘模型:Nemotron 3.5 Bolt 74B-A9B 和 Bolt E4B。
TL;DR
TL;DR(太长不看版)
- Size: 120B total parameters, 12B active (MoE). Context up to 1M tokens.
- Runs on: 1x DGX Spark, 1x H200, 2x H100 or 1x GB200 with the NVFP4 checkpoint. A BF16 checkpoint is also published.
- Performance: Strong reasoning, agentic coding and document vision. NVIDIA claims up to 5x higher single-GPU throughput than open peers.
- Best: 89.90 on GPQA-Diamond (NVFP4).
- Worst: 17.78 on tau3-bench banking (NVFP4).
- 规模:总共 120B 参数,12B 活跃参数(MoE)。上下文长度最高达 1M tokens。
- 运行环境:1x DGX Spark、1x H200、2x H100 或 1x GB200(使用 NVFP4 检查点)。同时也发布了 BF16 检查点。
- 性能:具备强大的推理、智能体编码和文档视觉能力。NVIDIA 声称其单 GPU 吞吐量比同类开源模型高出高达 5 倍。
- 最佳表现:在 GPQA-Diamond 上得分为 89.90(NVFP4)。
- 最差表现:在 tau3-bench banking 上得分为 17.78(NVFP4)。
What is Nemotron 3.5 Super VL?
什么是 Nemotron 3.5 Super VL?
It is an open multimodal agent model. Weights ship under the OpenMDW-1.1 license. NVIDIA also released open datasets, training recipes and RL environments in NeMo Gym.
它是一个开放的 multimodal 智能体模型。权重采用 OpenMDW-1.1 许可证发布。NVIDIA 还在 NeMo Gym 中发布了开放数据集、训练配方和强化学习环境。
The model card lists text, image and video inputs with text output. Video is limited to MP4 clips up to 2 minutes. Reasoning can be toggled on or off through the chat template.
模型卡片列出了文本、图像和视频输入以及文本输出。视频限制为最长 2 分钟的 MP4 片段。可以通过聊天模板开启或关闭推理功能。
How does the architecture work?
架构是如何工作的?
The backbone interleaves Mamba-2, mixture-of-experts and attention layers. Each MoE layer holds 512 routed experts. A C-RADIO vision encoder and a connector feed images, document pages and video frames into the language backbone.
骨干网络交错排列 Mamba-2、混合专家和注意力层。每个 MoE 层包含 512 个路由专家。C-RADIO 视觉编码器和连接器将图像、文档页面和视频帧输入到语言骨干网络中。
Speed comes from four techniques:
速度来自四种技术:
- Multi-token prediction: a native drafter proposes several tokens, verified in 1 forward pass.
- LiLiCorr: a speculative drafter for DGX Spark, up to 2.3x faster token generation.
- NVFP4 with quantization-aware distillation: 4-bit weights and activations on Blackwell, 4-bit weights with 16-bit activations elsewhere.
- Context reuse: prefix caching keeps both attention caches and Mamba states across agent turns.
- 多 token 预测:原生草稿生成器提出多个 token,在一次前向传递中进行验证。
- LiLiCorr:一种用于 DGX Spark 的推测性草稿生成器,token 生成速度最高快 2.3 倍。
- NVFP4 结合量化感知蒸馏:在 Blackwell 上使用 4-bit 权重和激活值,在其他地方使用 4-bit 权重和 16-bit 激活值。
- 上下文复用:前缀缓存技术在智能体交互轮次之间同时保留注意力缓存和 Mamba 状态。
The model was distilled from Nemotron 3 Ultra. NVIDIA team states that it trained on about 26 trillion tokens. Post-training used roughly 3.8 million samples plus RL across dozens of environments. Continued pretraining added 10B Indic tokens.
该模型是从 Nemotron 3 Ultra 蒸馏而来的。NVIDIA 团队表示,它在约 26 万亿个 token 上进行了训练。后训练使用了大约 380 万个样本,并在数十个环境中进行了强化学习。继续预训练添加了 10B 个 Indic token。
How does it perform on benchmarks?
它在基准测试中的表现如何?
All scores below are NVIDIA-reported, from the NVFP4 checkpoint on the model card.
以下所有分数均为 NVIDIA 报告的数据,来自模型卡片上的 NVFP4 检查点。
| Area | Benchmark | Score |
|---|---|---|
| Reasoning | GPQA-Diamond | 89.90 |
| Coding | SWE-bench Verified | 71.07 |
| Coding | Terminal-Bench 2.1 | 60.80 |
| Web agent | BrowseComp | 77.5 |
| Instruction following | IFBench | 71.50 |
| Charts | CharXiv reasoning | 79.88 |
| Documents | MMLongBench-Doc | 62.82 |
| Computer use | OSWorld-Verified | 52.3 |
| Long-horizon agent | tau3-bench banking | 17.78 |
| 领域 | 基准测试 | 得分 |
|---|---|---|
| 推理能力 | GPQA-Diamond | 89.90 |
| 编程能力 | SWE-bench Verified | 71.07 |
| 编程能力 | Terminal-Bench 2.1 | 60.80 |
| 网页智能体 | BrowseComp | 77.5 |
| 指令遵循能力 | IFBench | 71.50 |
| 图表分析 | CharXiv reasoning | 79.88 |
| 文档处理 | MMLongBench-Doc | 62.82 |
| 计算机操作 | OSWorld-Verified | 52.3 |
| 长周期智能体 | tau3-bench banking | 17.78 |
On agent harness evals, Super VL beats Nemotron 3 Ultra. It scores 63.44% on Hermes Bench versus 56.33%. On LangChain evals it scores 87.05% versus 84.3%. Ultra carries about 4.6x more total parameters.
在智能体工具包评估中,Super VL 优于 Nemotron 3 Ultra。它在 Hermes Bench 上得分为 63.44%,而 Nemotron 3 Ultra 为 56.33%。在 LangChain 评估中,其得分为 87.05%,对比值为 84.3%。Ultra 版本的总参数量约为前者的 4.6 倍。
NVIDIA team reports over 600 output tokens per second on Artificial Analysis agentic tasks via a Baseten endpoint. Its chart marks a 5x speed gap over Inkling Small.
NVIDIA 团队报告称,通过 Baseten 端点执行 Artificial Analysis 智能体任务时,每秒可输出超过 600 个 token。其图表显示,速度比 Inkling Small 快 5 倍。
更进一步:量化金融体系
看懂新闻只是起点——沿量化金融路径,把它变成能交付的工程能力