[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$f1vecv824xrz6r":3},{"slug":4,"category":5,"publishedAt":6,"titleZh":7,"titleEn":8,"summaryZh":9,"summaryEn":10,"models":11,"hardwares":21,"bodyZh":31,"bodyEn":32,"sourceName":33,"sourceUrl":34,"updatedAt":35},"nvidia-nemotron-35-super-vl-bolt-2026-10-24","NEWS","2026-10-24T14:00:00.000Z","NVIDIA 同日双发 Nemotron 3.5：120B-A12B 的 Super VL 与 6GB 边缘模型 Bolt","NVIDIA ships two Nemotron 3.5 models in one day: the 120B-A12B Super VL and a 6GB edge model, Bolt","10 月 23 日 NVIDIA 一次放出两个 Nemotron 3.5 模型。Super-VL-120B-A12B-NVFP4：Mamba-2 与 Transformer 混合 MoE + C-RADIO 视觉编码器，1M 上下文，NVFP4 权重实测 75.0GB，官方部署到 DGX Spark \u002F RTX PRO 6000 \u002F H200 \u002F GB200。Bolt-E4B-4GB：8B-A4B 的边缘小模型，Q4_K_M GGUF 仅 6.18GB，主打工具调用与设备侧语音助手。站内新增 103 条实测。","NVIDIA released two Nemotron 3.5 models on the same day, Oct 23. Super-VL-120B-A12B-NVFP4 mixes Mamba-2 and Transformer layers in a hybrid MoE with a C-RADIO vision encoder, carries a 1M context and measures 75.0GB of NVFP4 weights, and deploys officially on DGX Spark \u002F RTX PRO 6000 \u002F H200 \u002F GB200. Bolt-E4B-4GB is the 8B-A4B edge sibling, a 6.18GB Q4_K_M GGUF aimed at tool use and on-device voice assistants. The batch adds 103 new runs on-site.",[12,15,18],{"id":13,"name":14},67,"gemma-4-E4B-it",{"id":16,"name":17},96,"NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4",{"id":19,"name":20},97,"NVIDIA-Nemotron-3.5-Bolt-E4B-4GB",[22,25,28],{"id":23,"name":24},13,"NVIDIA RTX 5090",{"id":26,"name":27},54,"NVIDIA DGX Spark",{"id":29,"name":30},68,"NVIDIA RTX PRO 6000 Blackwell 96GB","# NVIDIA 同日双发 Nemotron 3.5：120B-A12B 的 Super VL 与 6GB 边缘模型 Bolt\n\n10 月 23 日，NVIDIA 一天之内放出两个 Nemotron 3.5 模型——一头一尾：一个是能吃视频、能做工具调用、1M 上下文的 120B 视觉语言大模型；另一个是 6GB 出头的边缘小模型，专攻设备侧语音助手。\n\n## Super-VL-120B-A12B-NVFP4\n\n**120B 总参 \u002F 12B 激活**的视觉语言 MoE，本次发布的是 **NVFP4 量化版**（由 BF16 版做量化感知蒸馏而来）。架构是交错的 **Mamba-2 + Transformer 混合注意力 + MoE**，配 **C-RADIO 视觉编码器**和视觉-语言连接器；上下文最长 **1,000,000 token**。\n\n几个工程细节值得记：\n\n- **原生 MTP 投机解码**：官方配 LiLiCorr 草稿模型，每步 7 个投机 token；vLLM 侧用 `--speculative_config.method dflash`\n- **推理档位可调**：`off` \u002F `on (medium)` \u002F `on`，走 chat template 控制，medium 档省 token 但可能掉准确率\n- **部署目标**：1× DGX Spark (GB10)、1× RTX PRO 6000、1× H200、2× H100、1× GB200；官方 vLLM 命令行还带 `--mamba-backend flashinfer`、`--video-pruning-rate 0.5`、`--media-io-kwargs` 视频 fps 2 \u002F 256 帧\n- **权重实测 75.0GB**（HF 仓库 88 个 safetensors 分片相加）\n\n量化几乎不掉点（NVFP4 \u002F BF16）：\n\n| 基准 | NVFP4 | BF16 |\n|---|---|---|\n| GPQA-Diamond | 89.90 | 89.52 |\n| Terminal-Bench 2.1 | **62.78** | 60.67 |\n| SWE-bench Verified | 71.07 | 71.10 |\n| SWE-bench Pro | 51.30 | 51.78 |\n| tau2-bench retail | 69.30 | 69.24 |\n| tau2-bench airline | 75.75 | 76.50 |\n| Video-MME v2 | 32.54 | 32.85 |\n| OSWorld-Verified | 52.3 | 51.14 |\n| AA-Omniscience index | 5.3 | 4.03 |\n| GDPval-AA v2 | 1206 | 1211 |\n\n许可为 OpenMDW-1.1，可商用。\n\n**站内口径**：已收录 **NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4**（id 96），**零实测记录**——原因和上次 RTX Spark 那篇一样是两本账：物理上 75GB 的 NVFP4 权重能塞进 96GB 的 PRO 6000，但站内是保守 padding 口径（120B × NVFP4 需要 180GB 以上的账），96GB 级显卡判不出 fit。真要跑，请按 NVIDIA 的部署清单来。\n\n## Bolt-E4B-4GB：6.18GB 的边缘小模型\n\n另一个方向：**NVIDIA-Nemotron-3.5-Bolt-E4B-4GB**——底座是 Google **Gemma-4-E4B-it**（Apache-2.0），NVIDIA 用 `nemotron-post-training-v3` 数据后训练，**8B 总参 \u002F 4B 有效参**，后训练数据截止 2026-05。\n\n- **Q4_K_M GGUF 实测 6.18GB**（仓库名里的「4GB」指 4B 有效参数量级，不是文件大小）\n- **llama.cpp 官方配方**：`ghcr.io\u002Fggml-org\u002Fllama.cpp:server-cuda13-b11459`，`-c 65536`、`--jinja`，另配 **DFlash 草稿模型**做投机解码（`--spec-type draft-dflash --spec-draft-n-max 7`）\n- 上下文输入\u002F输出各 **64K**，支持 **36 种语言**\n- chat template 控制 thinking 模式与 conversational 模式，thinking 预算可调\n- 定位明确：**本地语音与设备助手、交互人设、边缘嵌入式**\n\n官方 Q4_K_M 口径的基准（对照 Gemma-4-E4B-it 与 Qwen3.5-4B 的 Q4_K_M）：\n\n| 基准 | Bolt-E4B | Gemma-4-E4B-it | Qwen3.5-4B |\n|---|---|---|---|\n| MultiChallenge | **55.60** | 39.36 | 52.23 |\n| RoleBench | **71.21** | 60.51 | 52.35 |\n| RoleMRC | **79.12** | 75.24 | 62.61 |\n| ToolSandbox | **74.13** | 68.31 | 74.51 |\n| IHEval | **82.40** | 66.61 | 62.87 |\n| Tau2-Bench | 52.32 | 38.72 | **61.74** |\n| BFCL-v3 | 67.31 | 65.03 | **70.56** |\n| RULER (64K) | 85.76 | 86.14 | **95.37** |\n| LongMemEval | 69.92 | 67.95 | **76.60** |\n\n提升最明显的是工具调用与角色扮演两项（ToolSandbox、IHEval、RoleBench）；但长文本与长上下文（RULER \u002F LongMemEval \u002F Tau2-Bench）仍是 Qwen3.5-4B 更强——想本地跑长文档任务，Bolt 不是最优解。\n\n## 站内收录与实测\n\n- **NVIDIA-Nemotron-3.5-Bolt-E4B-4GB**（id 97）入库，新增 **26 条 L2 实测**（Q4_K_M GGUF + BF16），全池均值 142.3 tok\u002Fs：\n\n| 硬件 | 框架 \u002F 量化 | 实测 |\n|---|---|---|\n| H100 | Ollama Q4_K_M | 345.3 tok\u002Fs |\n| RTX 5090 | Ollama Q4_K_M \u002F vLLM BF16 | 279.8 \u002F 171.5 tok\u002Fs |\n| RTX PRO 6000 96GB | Ollama Q4_K_M \u002F vLLM BF16 | 264.8 \u002F 177.6 tok\u002Fs |\n| RTX 4090 | llama.cpp Q4_K_M | 160.6 tok\u002Fs（DFlash 接受率 0.54） |\n| M2 Ultra 80GB | llama.cpp Q4_K_M | 119.5 tok\u002Fs（接受率 0.66） |\n| RTX Spark（116 GiB Linux 池） | llama.cpp \u002F Ollama Q4_K_M | 80.7 \u002F 79.2 tok\u002Fs |\n\n- 本批共新增 **103 条实测**，另补齐三个老模型的硬件缺口：gemma-4-E4B-it（23 条，覆盖此前没跑过的 4090 \u002F 3090 \u002F 7900 XTX \u002F 5080 \u002F 4080S \u002F 3060 \u002F H100 \u002F H20 \u002F PRO 6000 \u002F M5 Ultra \u002F Arc B580 \u002F RTX Spark）、Ministral-3-8B-Instruct-2512（28 条）、iFlytek Spark-X2.5-4B（26 条）。","# NVIDIA ships two Nemotron 3.5 models in one day: the 120B-A12B Super VL and a 6GB edge model, Bolt\n\nOn Oct 23 NVIDIA released two Nemotron 3.5 models on the same day — one huge, one tiny: a 120B vision-language model that handles video, tool use and a 1M context, and a 6GB edge model aimed at on-device voice assistants.\n\n## Super-VL-120B-A12B-NVFP4\n\nA **120B total \u002F 12B active** vision-language MoE, released this time in **NVFP4** (quantization-aware distillation from the BF16 build). The architecture interleaves **Mamba-2 and Transformer attention layers in a hybrid MoE**, with a **C-RADIO vision encoder** and a vision-language connector; context runs to **1,000,000 tokens**.\n\nEngineering details worth noting:\n\n- **Native MTP speculative decoding**: an official LiLiCorr drafter, seven speculative tokens per step; on the vLLM side via `--speculative_config.method dflash`\n- **Switchable reasoning tier**: `off` \u002F `on (medium)` \u002F `on`, controlled through the chat template; medium saves tokens but can cost accuracy\n- **Deployment targets**: 1x DGX Spark (GB10), 1x RTX PRO 6000, 1x H200, 2x H100, 1x GB200; the official vLLM command line also carries `--mamba-backend flashinfer`, `--video-pruning-rate 0.5` and `--media-io-kwargs` with video fps 2 \u002F 256 frames\n- **Weights measure 75.0GB** (88 safetensors shards added up on the HF repo)\n\nQuantization costs almost nothing (NVFP4 \u002F BF16):\n\n| Benchmark | NVFP4 | BF16 |\n|---|---|---|\n| GPQA-Diamond | 89.90 | 89.52 |\n| Terminal-Bench 2.1 | **62.78** | 60.67 |\n| SWE-bench Verified | 71.07 | 71.10 |\n| SWE-bench Pro | 51.30 | 51.78 |\n| tau2-bench retail | 69.30 | 69.24 |\n| tau2-bench airline | 75.75 | 76.50 |\n| Video-MME v2 | 32.54 | 32.85 |\n| OSWorld-Verified | 52.3 | 51.14 |\n| AA-Omniscience index | 5.3 | 4.03 |\n| GDPval-AA v2 | 1206 | 1211 |\n\nThe license is OpenMDW-1.1, commercial use permitted.\n\n**On-site status**: **NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4** (id 96) is cataloged with **zero on-site runs** — for the same two-ledger reason as the earlier RTX Spark story: physically, 75GB of NVFP4 weights fit a 96GB RTX PRO 6000, but the site runs a deliberately padded accounting where 120B at NVFP4 books above 180GB, so 96GB-class cards get no fit verdict. If you want to run it, follow NVIDIA's deployment matrix.\n\n## Bolt-E4B-4GB: a 6.18GB edge model\n\nThe other direction: **NVIDIA-Nemotron-3.5-Bolt-E4B-4GB** is built on Google's **Gemma-4-E4B-it** (Apache-2.0) and post-trained by NVIDIA on `nemotron-post-training-v3`, with **8B total \u002F 4B effective** parameters and a May 2026 post-training cutoff.\n\n- The **Q4_K_M GGUF measures 6.18GB** (the \"4GB\" in the repo name refers to the 4B effective-parameter class, not the file size)\n- **Official llama.cpp recipe**: `ghcr.io\u002Fggml-org\u002Fllama.cpp:server-cuda13-b11459` with `-c 65536` and `--jinja`, plus a **DFlash draft model** for speculative decoding (`--spec-type draft-dflash --spec-draft-n-max 7`)\n- 64K context in and out, **36 languages**\n- The chat template controls thinking mode and conversational mode, with a tunable thinking budget\n- Positioning is explicit: **local voice and device assistants, interactive personas, embedded edge**\n\nOfficial Q4_K_M benchmarks (against Gemma-4-E4B-it and Qwen3.5-4B at Q4_K_M):\n\n| Benchmark | Bolt-E4B | Gemma-4-E4B-it | Qwen3.5-4B |\n|---|---|---|---|\n| MultiChallenge | **55.60** | 39.36 | 52.23 |\n| RoleBench | **71.21** | 60.51 | 52.35 |\n| RoleMRC | **79.12** | 75.24 | 62.61 |\n| ToolSandbox | **74.13** | 68.31 | 74.51 |\n| IHEval | **82.40** | 66.61 | 62.87 |\n| Tau2-Bench | 52.32 | 38.72 | **61.74** |\n| BFCL-v3 | 67.31 | 65.03 | **70.56** |\n| RULER (64K) | 85.76 | 86.14 | **95.37** |\n| LongMemEval | 69.92 | 67.95 | **76.60** |\n\nThe clearest gains are in tool use and role play (ToolSandbox, IHEval, RoleBench); on long documents and long context (RULER \u002F LongMemEval \u002F Tau2-Bench) Qwen3.5-4B still leads — if your local workload is long-document work, Bolt is not the best pick.\n\n## On-site coverage\n\n- **NVIDIA-Nemotron-3.5-Bolt-E4B-4GB** (id 97) is cataloged with **26 new L2 records** (Q4_K_M GGUF + BF16), averaging 142.3 tok\u002Fs across the pool:\n\n| Hardware | Framework \u002F quant | Measured |\n|---|---|---|\n| H100 | Ollama Q4_K_M | 345.3 tok\u002Fs |\n| RTX 5090 | Ollama Q4_K_M \u002F vLLM BF16 | 279.8 \u002F 171.5 tok\u002Fs |\n| RTX PRO 6000 96GB | Ollama Q4_K_M \u002F vLLM BF16 | 264.8 \u002F 177.6 tok\u002Fs |\n| RTX 4090 | llama.cpp Q4_K_M | 160.6 tok\u002Fs (DFlash acceptance 0.54) |\n| M2 Ultra 80GB | llama.cpp Q4_K_M | 119.5 tok\u002Fs (acceptance 0.66) |\n| RTX Spark (116 GiB Linux pool) | llama.cpp \u002F Ollama Q4_K_M | 80.7 \u002F 79.2 tok\u002Fs |\n\n- The batch adds **103 runs** in total, also closing hardware gaps for three older models: gemma-4-E4B-it (23 runs on hardware it had never touched — 4090, 3090, 7900 XTX, 5080, 4080S, 3060, H100, H20, RTX PRO 6000, M5 Ultra, Arc B580, RTX Spark), Ministral-3-8B-Instruct-2512 (28 runs) and iFlytek Spark-X2.5-4B (26 runs).","araoai","https:\u002F\u002Faraoai.com\u002Fnews\u002Fnvidia-nemotron-35-super-vl-bolt-2026-10-24","2026-10-24T01:46:21.427Z"]