[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-nemotron-3-5-super-vl-open-120b-moe":3,"topics-all":41,"news-related-b44fb160-9027-44cd-b355-a66869004abc":60},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":27,"news_slug":34,"published_at":35,"created_at":36,"modified_at":37,"is_published":38,"publish_type":39,"image_url":14,"view_count":40},"b44fb160-9027-44cd-b355-a66869004abc","Nemotron 3.5 Super VL 开源:120B 总参只激活 12B","英伟达 10 月 15 日开源多模态 MoE Nemotron 3.5 Super VL:120B 总参数、激活仅 12B,最长 100 万 token 上下文,文本、图像、文档、视频理解加工具调用装进一个检查点。官方称单卡 Agent 吞吐为同级开放模型 5 倍,Scale AI 后训练后推理成本降 40 倍。","长时间运行的企业 Agent 正在接手越来越复杂的流程:一个请求可能要读一份多页文档、跨来源核对数据、遵守公司政策、再调用好几个工具。过去这类任务通常要“文本模型 + 视觉模型”接力,上下文在两套栈之间来回倒手。10 月 15 日,英伟达开放了 Nemotron 3.5 Super VL,把接力赛改成一个人跑:120B 参数的多模态混合专家(MoE)模型,推理、指令跟随、工具调用,加上文本、图像、文档、视频理解,全部装进同一个检查点([NVIDIA 技术博客](https:\u002F\u002Fdeveloper.nvidia.com\u002Fblog\u002Fnvidia-nemotron-3-5-super-vl-powers-specialized-agents-with-multimodal-reasoning))。\n\n## 一个模型,两套栈的活全包\n\nSuper VL 的视觉训练覆盖图表、表格、OCR、多页文档和视频,官方给出的定位是:文档智能场景里,同一个模型既能读懂一张图表,又能关联文档另一处的表格,再把两边总结到一起。图表推理基准 CharXiv 上它拿到 80.8%,对比上一代 Nemotron 3 Nano Omni 的 63%;配合 Python 代码解释还能到 84%。在 OCRBench v2 排行榜上,它也是排名靠前的开放模型之一。对做 Agent 产品的团队,这意味着不必再单独部署和维护一个视觉模型。\n\n## 120B 的容量,12B 的账单\n\n架构上,Super VL 是 120B 总参数、每 token 仅激活 12B 的 MoE,从更大的 Nemotron 3 Ultra 蒸馏而来,蒸馏阶段用了 SkewReverse 损失,后训练阶段用多教师在线策略蒸馏(MOPD)让文本和视觉教师同时给生成结果打分。根据 [Modal 的模型卡](https:\u002F\u002Fmodal.com\u002Flibrary\u002Fnvidia\u002Fnemotron-3-5-super-vl-120b-a12b-nvfp4),它采用混合 Mamba-Transformer 架构和 C-RADIOv4 视觉编码器,支持最长 100 万 token 上下文,并内置多 token 预测。官方口径的效率数字更激进:单 GPU 上 Agent 任务吞吐最高达到同级别开放模型的 5 倍,在 Baseten 端点上跑 Artificial Analysis Intelligence Index 任务输出速度超过每秒 600 token;一万道编码任务的整体完成时间也比同级别开放模型快 4 倍。\n\n## 推理优化四件套\n\n真正决定 always-on Agent 成本的是推理侧。Super VL 配了四件东西:LiLiCorr 开源投机解码,整份草稿联合评估 token 兼容性,DGX Spark 上生成提速最高 2.3 倍;NVFP4 量化加量化感知蒸馏(QAD),量化后的模型在训练中对齐 BF16 原版输出,KV 缓存也能存成 NVFP4 且不增加推理开销;多 token 预测(MTP)让模型一次前向验证多个候选 token;以及跨 Agent 轮次的上下文复用,同时保留注意力缓存和 Mamba 循环状态。部署侧,用 NVIDIA Dynamo 的 KV 感知路由,16 块 B200 在每用户 50 token\u002F秒的交互速率下可服务 8 倍用户,首 token 延迟降低 85%。\n\n## 企业后训练的账本\n\n开放权重的意义在后训练。Scale AI 把它后训练用于 Skydance 的自动供应商信贷处理,官方预计推理成本下降 40 倍、延迟下降 5 倍;金融平台 Ramp 用它把收据图片转成结构化 JSON,预期吞吐提升最高 50%、延迟降 10 倍;AI21 与 Alice 针对威胁情报和漏洞分析后训练,准确率提升 41%。官方展示的领域增益从文档智能 +15% 到医疗保险理赔 +44% 不等。基准上,Hermes Bench 63.44%、LangChain 评测 87.05%,都超过了更大的 Nemotron 3 Ultra。\n\n## 开放的完整拼图,和它的生意逻辑\n\n这次发布不只是丢一个检查点:配套开源的还有 Agent 工具使用、金融推理、长上下文等数据集(其中视觉语言数据集 40 个),以及与 IIT Madras 的 AI4Bharat 合作的 2.8T token 印度语言数据集,训练时先注入了 10B token 的 12 种语言子集。边缘侧还有两个伴生模型:Bolt 74B-A9B 面向 24GB 内存的本地设备,现已可用;4GB 的 Bolt E4B 主打语音助手,10 月 23 日跟上([The Decoder](https:\u002F\u002Fthe-decoder.com\u002Fnvidias-nemotron-3-5-super-vl-is-an-open-model-built-to-run-enterprise-ai-agents-nonstop))。The Decoder 的观察值得记住:模型完全开放,但整条优化链路都为英伟达自家硬件调校——想榨干性能,最后还是会买它的卡。\n\n所以呢?Agent 基建的竞争正在从“谁的模型跑分高”转向“谁能在单卡上把 always-on Agent 跑便宜”。英伟达用开放权重加全套数据、配方和部署工具,把门槛降到任何一个团队都能 fork 起步——代价是默认走进它的硬件栈。对开发者,这是一笔划算的买卖;对其他云厂商,这是教科书级的生态卡位。","https:\u002F\u002Fdeveloper.nvidia.com\u002Fblog\u002Fnvidia-nemotron-3-5-super-vl-powers-specialized-agents-with-multimodal-reasoning","474eef8c-e0c3-46cf-adee-c089558220f9",[11,15,18,21,24],{"id":12,"name":13,"slug":13,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"id":19,"name":20,"slug":20,"description":14,"color":14},"d11f0044-8aef-487c-bebe-89ce4683a4a3","moe",{"id":22,"name":23,"slug":23,"description":14,"color":14},"499f4b56-819d-49a3-9609-33e775143b86","multimodal",{"id":25,"name":26,"slug":26,"description":14,"color":14},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[28],{"id":29,"lang":30,"title":31,"summary":32,"content":33},"04ca47eb-6f29-40c9-bb3f-61a1203e6b66","en","Nemotron 3.5 Super VL: NVIDIA's open 120B MoE, 12B active","NVIDIA open-sources Nemotron 3.5 Super VL: a 120B multimodal MoE with 12B active and 1M context for agents. Claims 5x throughput; Scale AI cut cost 40x.","Always-on enterprise agents keep taking on longer workflows: a single request may read a multi-page document, cross-check sources, follow company policy, and call several tools. That used to mean a text model plus a separate vision-language model, passing context back and forth. On October 15, NVIDIA open-sourced Nemotron 3.5 Super VL and collapsed the relay into one runner: a 120B-parameter multimodal mixture-of-experts model that packs reasoning, instruction following, tool use, and text, image, document, and video understanding into a single checkpoint ([NVIDIA Technical Blog](https:\u002F\u002Fdeveloper.nvidia.com\u002Fblog\u002Fnvidia-nemotron-3-5-super-vl-powers-specialized-agents-with-multimodal-reasoning)).\n\n## One model, two stacks' worth of work\n\nSuper VL's visual training covers charts, tables, OCR, multipage documents, and video. NVIDIA's framing: in document intelligence, the same model can interpret a chart, connect it with a table elsewhere in the document, and summarize both. On CharXiv, a chart-reasoning benchmark, it scores 80.8% versus 63% for the previous Nemotron 3 Nano Omni, and 84% when paired with Python code interpretation. It also ranks among the top open models on the OCRBench v2 leaderboard. For teams building agents, that removes the need to deploy and maintain a separate vision model.\n\n## 120B of capacity, a 12B bill\n\nArchitecturally, Super VL is a MoE with 120B total parameters and only 12B active per token. It was distilled from the larger Nemotron 3 Ultra using the SkewReverse loss, with multi-teacher on-policy distillation (MOPD) in post-training where text and vision teachers score generated responses. According to [Modal's model card](https:\u002F\u002Fmodal.com\u002Flibrary\u002Fnvidia\u002Fnemotron-3-5-super-vl-120b-a12b-nvfp4), it uses a hybrid Mamba-Transformer architecture with the C-RADIOv4 vision encoder, supports up to 1 million tokens of context, and has built-in multi-token prediction. NVIDIA's efficiency claims are aggressive: up to 5x higher throughput for agentic tasks on a single GPU versus other open models in its size category, over 600 output tokens per second on the Baseten endpoint running Artificial Analysis Intelligence Index tasks, and 4x faster completion of 10,000 coding tasks.\n\n## Four tricks on the inference side\n\nWhat really determines the cost of an always-on agent is inference. Super VL ships with four pieces: LiLiCorr, an open speculative-decoding technique that jointly evaluates token compatibility across an entire draft, delivering up to 2.3x faster generation on DGX Spark; NVFP4 quantization with quantization-aware distillation (QAD), where the quantized model learns to match the BF16 original during training and the KV cache can also be stored in NVFP4 with no inference overhead; multi-token prediction (MTP), verifying several candidate tokens per forward pass; and context reuse across agent turns that preserves both attention caches and Mamba recurrent states. On the serving side, NVIDIA Dynamo's KV-aware routing serves up to 8x as many users on 16 B200 GPUs at 50 tokens per second per user, with 85% lower first-token latency.\n\n## The enterprise post-training ledger\n\nOpen weights matter because of post-training. Scale AI post-trained the model for automated vendor credit processing at Skydance; NVIDIA projects 40x lower inference cost and 5x lower latency. fintech platform Ramp converts receipt images into structured JSON and expects up to 50% higher throughput with 10x lower latency. AI21 and Alice post-trained it for threat intelligence and vulnerability analysis, improving accuracy by 41%. NVIDIA's shown domain gains range from +15% on document intelligence to +44% on medical insurance claims. On benchmarks, it scores 63.44% on Hermes Bench and 87.05% on LangChain evals, beating the larger Nemotron 3 Ultra on both.\n\n## The full open puzzle — and the business logic\n\nThis release is more than a dropped checkpoint: alongside the weights, NVIDIA open-sourced datasets for agentic tool use, finance reasoning, and long-context work (including more than 40 vision-language datasets), plus a 2.8T-token Indian-language dataset built with IIT Madras's AI4Bharat lab, after injecting a 10B-token, 12-language subset during training. Two companion models target the edge: Bolt 74B-A9B, for local machines with 24GB of memory, is available now, while the 4GB Bolt E4B for voice assistants follows on October 23 ([The Decoder](https:\u002F\u002Fthe-decoder.com\u002Fnvidias-nemotron-3-5-super-vl-is-an-open-model-built-to-run-enterprise-ai-agents-nonstop)). The Decoder's observation is worth remembering: the model is fully open, but the whole optimization chain is tuned for NVIDIA's own hardware — squeeze the performance and you end up buying its chips.\n\nSo what? The agent-infrastructure race is shifting from \"whose model scores highest\" to \"who can run an always-on agent cheaply on a single GPU.\" NVIDIA pairs open weights with datasets, recipes, and deployment tooling, lowering the barrier until any team can fork and start — at the cost of walking into its hardware stack by default. For developers, that's a good deal; for rival clouds, it's textbook ecosystem positioning.","nemotron-3-5-super-vl-open-120b-moe","2026-10-20T17:11:56Z","2026-10-20T17:12:10.390991Z","2026-10-20T17:12:10.390999Z",true,"agent",177,[42,51],{"slug":43,"tag_slug":43,"title_zh":44,"title_en":45,"intro_zh":46,"intro_en":47,"id":48,"is_active":38,"created_at":49,"modified_at":50},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":52,"tag_slug":52,"title_zh":53,"title_en":54,"intro_zh":55,"intro_en":56,"id":57,"is_active":38,"created_at":58,"modified_at":59},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":61},[62,67,72,77,82,87],{"id":63,"title":64,"news_slug":65,"published_at":66},"ac666c4e-4756-43d8-840a-4eacbf88f804","Meta 开源 816B MoE 模型 Muse Spark Open,跑在两台 Mac Studio 上","meta-muse-spark-open-816b-mac-studio","2026-10-17T03:00:00+00:00",{"id":68,"title":69,"news_slug":70,"published_at":71},"1b6ec126-cba1-47da-870b-da19cc523763","Meta 开源 816B MoE 模型 Muse Spark Open:1M 上下文,两台 Mac Studio 跑得动","meta-muse-spark-open-816b-moe-1m-mac-studio","2026-10-15T11:00:00+00:00",{"id":73,"title":74,"news_slug":75,"published_at":76},"faf5c85e-ed83-432a-8b55-80c9356862c5","腾讯优图开源5B全模态解析:文档榜96.96超GPT-5.2","youtu-parsing-omni-5b","2026-10-19T19:09:27+00:00",{"id":78,"title":79,"news_slug":80,"published_at":81},"bbfcbee4-8d92-4078-bb1e-eaf4d1c1e146","Beam 权重落地:501B MoE 三精度齐发","beam-1-0-weights-open-agent-stack","2026-10-18T21:10:09+00:00",{"id":83,"title":84,"news_slug":85,"published_at":86},"d529114c-bfd7-4efb-9af4-444d3065b397","Step 5 Preview 开源,2.7 美元\u002F百万输出","stepfun-step-5-preview-600b-moe","2026-10-17T21:12:04+00:00",{"id":88,"title":89,"news_slug":90,"published_at":91},"cee60a45-f212-4a78-83dd-0226c07b3460","腾讯开源记忆写入模型 Hy-Memory:30B MoE 只激活 3B,四框架平均分超 Opus 5","tencent-hy-memory-30b-memory-writer","2026-10-17T15:14:21+00:00"]