[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-nemotron-35-lightning-30b-a3b-open-release":3,"news-related-de2cceb2-7d39-4a5f-844e-5a3144667f49":38},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"de2cceb2-7d39-4a5f-844e-5a3144667f49","Nemotron 3.5 Lightning 开源：30B 总参 3B 激活的混合 MoE，直接用 NVFP4 配方预训练","NVIDIA 于 8 月 11 日开源 Nemotron 3.5 Lightning（30B 总参数、3B 激活），采用 Mamba-2 与 MoE 及 Attention 混合架构，上下文最高 1M token，OpenMDW v1.1 许可商用可用。模型用 NVFP4 配方预训练超过 20 万亿 token，经 MTP、SFT、GRPO 强化学习与训练后量化五阶段完成，BF16 与 NVFP4 版跑分差距普遍在 1 个点以内。","8 月 11 日，NVIDIA 把 Nemotron 3.5 Lightning（30B-A3B）放上 Hugging Face 和 Build.NVIDIA.com，OpenMDW v1.1 许可，商用开放。Nemotron 家族的定位一直很明确——开放权重、开放训练数据、开放配方，Lightning 这一档瞄准的则是\"sub-agent 干活模型\"：长跑自主智能体、子智能体部署和智能体工作流。Build.NVIDIA.com 页面给它的标语是\"最快的 30B A3B MoE 模型\"——这是 NVIDIA 自家的说法，但配合 3B 激活参数的设计意图，方向感是清楚的：这类模型不追求单步推理登顶，而是要便宜、稳定地扛住高频调用。\n\n## 架构：三种组件混着搭\n\nLightning 3.5 用的是混合 MoE 架构：Mamba-2 层、MoE 层和部分 Attention 层交错排布，总参数 30B、每 token 激活 3B，上下文窗口最高 1M token。支持英语和编程语言，外加西班牙语、法语、德语、意大利语、日语。推荐采样参数直接写在模型卡里：Temperature 1.0、Top_P 0.95。\n\n## 训练：五个阶段，从 NVFP4 预训练到 PTQ\n\n模型卡的训练方法学分五段，每段都有可查证的细节：\n\n**阶段一，预训练**：超过 20 万亿 token，用 NVFP4 配方训练——注意这是\"用低精度格式预训练\"，不是训练后量化。预训练数据截止 2025 年 9 月，后训练数据截止 2026 年 5 月。软件栈是 Megatron-LM。\n\n**阶段二，MTP 继续预训练**：给模型加 Multi-Token Prediction 层，让模型同时预测多个未来 token，提供更丰富的训练信号。\n\n**阶段三，SFT**：合成代码、数学、科学、工具调用、指令遵循、结构化输出数据，外加长程检索和多文档聚合任务。\n\n**阶段四，强化学习**：用 GRPO（Group Relative Policy Optimization）跨数学、代码、科学、指令遵循、多步工具使用、多轮对话、结构输出环境做多环境 RL，采用解耦训练与推理的异步架构，并利用 MTP 加速 rollout 生成。\n\n**阶段五，PTQ**：用 NVIDIA Model Optimizer 做训练后量化——\"Four Over Six NVFP4\"静态 MSE 校准变体，路由专家和共享专家走 W4A16，Mamba 投影层和 KV cache 走 FP8 逐张量动态缩放，校准用 1000 个 32k 长度样本。\n\n值得一提的是后训练合成数据的透明度：模型卡列出了教师模型清单，包括 DeepSeek-V4-Pro、Nemotron 5.5、GPT-5.5、GLM-5、gpt-oss-120b 等——谁给谁蒸馏了一目了然。\n\n## 跑分：BF16 与 NVFP4 几乎打平\n\n官方评估表同时给了 BF16 和 NVFP4 两列，这比单列数字有价值得多：\n\n| 基准 | BF16 | NVFP4 |\n|---|---|---|\n| MMLU Pro | 81.94 | 81.62 |\n| GPQA Diamond（无工具） | 75.44 | 75.57 |\n| SWE-bench Verified | 51.56 | 52.80 |\n| Terminal-Bench 2.1 | 24.58 | 23.46 |\n| BrowseComp | 36.97 | 36.81 |\n| AA-LCR（长上下文） | 52.00 | 49.19 |\n\n大多数基准上 NVFP4 与 BF16 差距在 1 个点以内，SWE-bench Verified 甚至反超（52.80 对 51.56）；长上下文的 AA-LCR 掉得多一些（52.00 到 49.19）。评估配方、容器、提示词和评分设置全部发布在 NeMo Gym 仓库，可复现。SWE-bench 和 Terminal-Bench 用的是 NeMo Evaluator。\n\n## 部署侧\n\n硬件兼容三档：Blackwell、Hopper（NVFP4\u002FW4A16）、Ampere（W4A16）。推理引擎 Dynamo + vLLM，测试硬件 H100。另外 NGC 目录还有一个 BF16 全精度版本，官方说明它主要面向定制化和后训练，而非直接生产推理——生产路径默认走量化版。\n\n## 所以呢\n\nNemotron 3.5 Lightning 最值得记住的不是跑分，而是两件事：一是 NVFP4 直接进预训练配方，量化和训练不再是两个割裂的工序；二是模型卡把 20T 预训练语料、合成数据教师模型、RL 环境、量化校准参数全部摊开——在\"开放\"这件事上，NVIDIA 已经把标准抬到了\"开放到可复现\"的层级。做智能体系统的开发者，可以把它当成一份可直接拿来做 sub-agent 底座的公开工程参考。\n\n（信息源：https:\u002F\u002Fbuild.nvidia.com\u002Fnvidia\u002Fnemotron-3.5-lightning-30b-a3b\u002Fmodelcard ）","https:\u002F\u002Fbuild.nvidia.com\u002Fnvidia\u002Fnemotron-3.5-lightning-30b-a3b\u002Fmodelcard","474eef8c-e0c3-46cf-adee-c089558220f9",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"d11f0044-8aef-487c-bebe-89ce4683a4a3","moe",{"id":19,"name":20,"slug":20,"description":14,"color":14},"8dac812d-3839-4abe-a855-5f56ec9515fd","nvidia",{"id":22,"name":23,"slug":23,"description":14,"color":14},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"b714f55a-80c1-4163-9332-5ef763198555","en","Nemotron 3.5 Lightning: 30B MoE pretrained in NVFP4","NVIDIA open-sourced Nemotron 3.5 Lightning on August 11 — 30B total parameters with 3B active per token — using a hybrid Mamba-2 + MoE + Attention architecture with up to 1M tokens of context, commercially usable under the OpenMDW v1.1 license. The model was pre-trained on over 20 trillion tokens with an NVFP4 recipe, going through five stages including MTP, SFT, GRPO reinforcement learning, and post-training quantization; BF16 and NVFP4 benchmark scores stay within roughly 1 point of each other.","On August 11, NVIDIA put Nemotron 3.5 Lightning (30B-A3B) on Hugging Face and Build.NVIDIA.com under the OpenMDW v1.1 license, cleared for commercial use. The Nemotron family has always had a clear positioning — open weights, open training data, open recipes — and the Lightning tier targets the \"sub-agent workhorse\": long-running autonomous agents, sub-agent deployments, and agentic workflows. The Build.NVIDIA.com page tags it as \"the fastest 30B A3B MoE model\" — that's NVIDIA's own claim, but combined with the 3B-active-parameter design intent, the direction is clear: models in this class don't chase single-step reasoning leaderboard wins; they aim to absorb high-frequency calls cheaply and reliably.\n\n## Architecture: Three Components Interleaved\n\nLightning 3.5 uses a hybrid Mixture-of-Experts architecture: Mamba-2 layers, MoE layers, and select Attention layers are interleaved, with 30B total parameters and 3B active per token, plus a context window of up to 1M tokens. It supports English and coding languages, along with Spanish, French, German, Italian, and Japanese. Recommended sampling parameters are stated right in the model card: Temperature 1.0, Top_P 0.95.\n\n## Training: Five Stages, from NVFP4 Pre-training to PTQ\n\nThe model card's training methodology section breaks down into five stages, each with verifiable details:\n\n**Stage 1, pre-training**: more than 20 trillion tokens, trained with an NVFP4 recipe — note this is \"pre-training in a low-precision format,\" not post-training quantization. Pre-training data cutoff is September 2025; post-training data cutoff is May 2026. The software stack is Megatron-LM.\n\n**Stage 2, continued pre-training for MTP**: the model gains Multi-Token Prediction layers that predict multiple future tokens simultaneously, providing richer training signals.\n\n**Stage 3, SFT**: synthetic code, math, science, tool-calling, instruction-following, and structured-output data, plus tasks designed for long-range retrieval and multi-document aggregation.\n\n**Stage 4, reinforcement learning**: multi-environment RL using GRPO (Group Relative Policy Optimization) across math, code, science, instruction following, multi-step tool use, multi-turn conversation, and structured-output environments, with an asynchronous architecture that decouples training from inference and leverages MTP to accelerate rollout generation.\n\n**Stage 5, PTQ**: post-training quantization with NVIDIA Model Optimizer — a \"Four Over Six NVFP4\" variant of static MSE calibration, W4A16 on routed and shared experts, FP8 per-tensor dynamic scales on Mamba projections and KV cache, calibrated with 1000 samples at 32k token length.\n\nOne thing worth highlighting is the transparency of post-training synthetic data: the model card lists the teacher models — including DeepSeek-V4-Pro, Nemotron 5.5, GPT-5.5, GLM-5, and gpt-oss-120b — so who distilled from whom is fully visible.\n\n## Benchmarks: BF16 and NVFP4 Nearly Neck and Neck\n\nThe official evaluation table gives both BF16 and NVFP4 columns, which is far more valuable than a single column of numbers:\n\n| Benchmark | BF16 | NVFP4 |\n|---|---|---|\n| MMLU Pro | 81.94 | 81.62 |\n| GPQA Diamond (no tools) | 75.44 | 75.57 |\n| SWE-bench Verified | 51.56 | 52.80 |\n| Terminal-Bench 2.1 | 24.58 | 23.46 |\n| BrowseComp | 36.97 | 36.81 |\n| AA-LCR (long context) | 52.00 | 49.19 |\n\nOn most benchmarks NVFP4 lands within 1 point of BF16, and SWE-bench Verified even comes out ahead (52.80 vs 51.56); long-context AA-LCR drops a bit more (52.00 to 49.19). Evaluation recipes, containers, prompts, and scoring configurations are all published in the NeMo Gym repository for reproducibility. SWE-bench and Terminal-Bench used NeMo Evaluator.\n\n## On the Deployment Side\n\nHardware compatibility spans three tiers: Blackwell, Hopper (NVFP4\u002FW4A16), and Ampere (W4A16). Inference engine: Dynamo + vLLM, tested on H100. The NGC catalog also hosts a BF16 full-precision version, which the official notes say is intended primarily for customization and post-training rather than direct production inference — the production path defaults to the quantized build.\n\n## So What\n\nThe most memorable things about Nemotron 3.5 Lightning aren't the benchmark numbers, but two structural facts. First, NVFP4 goes directly into the pre-training recipe — quantization and training are no longer two disconnected processes. Second, the model card lays everything bare: the 20T pre-training corpus, synthetic-data teacher models, RL environments, quantization calibration parameters. On the \"openness\" front, NVIDIA has raised the bar to \"open enough to reproduce.\" Developers building agentic systems can treat it as a public engineering reference ready to use as a sub-agent foundation.\n\n(Source: https:\u002F\u002Fbuild.nvidia.com\u002Fnvidia\u002Fnemotron-3.5-lightning-30b-a3b\u002Fmodelcard )","nemotron-35-lightning-30b-a3b-open-release","2026-08-16T15:00:00Z","2026-08-15T19:07:26.146000Z","2026-08-15T19:07:26.146016Z",true,"agent",169,{"items":39},[40,45,50,55,60,65],{"id":41,"title":42,"news_slug":43,"published_at":44},"804ab59a-a8d6-4b61-bf74-8f6f2bdae83c","智谱把 Flash 做成一件正经事:一次说清 GLM-5.3-Flash 的架构和 benchmark 真相","glm-5-3-flash-hybrid-attention-architecture","2026-08-27T08:00:00+00:00",{"id":46,"title":47,"news_slug":48,"published_at":49},"7958a2f1-028c-4b4e-b134-0d5de9afc1c1","Motif 3 收官:韩国 314B MoE 改用 MIT 许可,从零起步架构首次面向商用","motif-3-mit-license-sovereign-ai","2026-08-24T00:00:00+00:00",{"id":51,"title":52,"news_slug":53,"published_at":54},"96b989b7-992b-424e-a8c1-1568760150c1","小红书开源 dots3-note:280B MoE 多模态、512K 上下文,Apache 2.0 直接放行","dots3-note-preview-280b-open-weights","2026-08-18T23:10:00+00:00",{"id":56,"title":57,"news_slug":58,"published_at":59},"491f4904-c854-4925-b3e3-e34b8afd5e50","KDA+MLA 混合栈下沉到 1.3B 激活:Ling-3.0-tiny 把 MoE 端侧化,INT4 跑出 115 tok\u002Fs","ling-3-tiny-kda-mla-edge-deployment","2026-08-18T00:00:00+00:00",{"id":61,"title":62,"news_slug":63,"published_at":64},"bb42bbdc-55a5-422c-ae68-4d103639bfd2","Ling-3.0-tiny：M4 Pro 实测 86-90 tokens\u002Fs","ling-3-0-tiny-edge-deployment-ledger","2026-08-17T23:30:00+00:00",{"id":66,"title":67,"news_slug":68,"published_at":69},"41c8976c-d775-4603-aa01-693c57b7b0bd","NVIDIA Nemotron 3 Embed 登顶 RTEB：把 8B 旗舰检索能力蒸馏进 1B 部署款","nvidia-nemotron-3-embed","2026-07-16T18:00:00+00:00"]