[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$fva30iptcd4ps":3,"$f3a1usmul145x4":30,"$fc81vg4g18y82":32,"$f3boz9dhx7w5cm":36,"$f8ypa021er9qz":52,"$f2y9v1wgiksu27":54},{"id":4,"name":5,"publisher":6,"paramsB":7,"useCase":8,"licenseStatus":9,"licenseUrl":10,"sourceUrl":11,"descriptionZh":12,"descriptionEn":13,"imageUrl":14,"downloadUrls":15,"stats":21,"variants":24,"derivations":25,"measuredHardware":28,"updatedAt":29},96,"NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4","NVIDIA",120,"REASONING","OPEN_WEIGHTS","https:\u002F\u002Fopenmdw.ai\u002Flicense\u002F1-1\u002F","https:\u002F\u002Fhuggingface.co\u002Fnvidia\u002FNVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4","NVIDIA 于 2026-10-23 发布的 Nemotron 3.5 Super VL 量化版（NVFP4，由 BF16 版做量化感知蒸馏）：120B 总参 \u002F 12B 激活的视觉-语言 MoE。架构为交错的 Mamba-2 + Transformer 混合注意力 + MoE，配 C-RADIO 视觉编码器与视觉-语言连接器；上下文最长 **1,000,000 token**，原生 Multi-Token Prediction（MTP）投机解码（官方配 LiLiCorr 草稿模型，每步 7 个投机 token）。推理档位可调 off \u002F on (medium) \u002F on，经 chat template 控制；部署目标为 1× DGX Spark (GB10)、1× RTX PRO 6000、1× H200、2× H100 或 1× GB200，需 vLLM（官方给 nightly 镜像 + `--speculative_config.method dflash` 等参数，mamba-backend flashinfer、video-pruning-rate 0.5）。NVFP4 权重实测合计 **75.0GB**（HF 仓库 88 个 safetensors 分片）。基准（NVFP4 \u002F BF16）：GPQA-Diamond 89.90 \u002F 89.52、Terminal-Bench 2.1 62.78 \u002F 60.67、SWE-bench Verified 71.07 \u002F 71.10、SWE-bench Pro 51.30 \u002F 51.78、tau2-bench retail 69.30 \u002F 69.24、tau2-bench airline 75.75 \u002F 76.50、Video-MME v2 32.54 \u002F 32.85、OSWorld-Verified 52.3 \u002F 51.14、AA-Omniscience index 5.3 \u002F 4.03、GDPval-AA v2 1206 \u002F 1211。许可为 OpenMDW-1.1（可商用）。**站内按保守显存口径零实测收录**：120B 级 NVFP4 在站内 padding 公式下需约 180GB 以上，96GB 级显卡不判 fit。","NVIDIA's quantized release of Nemotron 3.5 Super VL (2026-10-23, NVFP4 with quantization-aware distillation from the BF16 model): a 120B-total \u002F 12B-active vision-language MoE. 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 up to **1,000,000 tokens**, and native Multi-Token Prediction (MTP) enables speculative decoding (official LiLiCorr drafter, seven speculative tokens per step). Reasoning level is switchable between off \u002F on (medium) \u002F on through the chat template; deployment targets are 1x DGX Spark (GB10), 1x RTX PRO 6000, 1x H200, 2x H100 or 1x GB200, running on vLLM (official nightly image with --speculative_config.method dflash, mamba-backend flashinfer, video-pruning-rate 0.5). The NVFP4 weights measure **75.0GB** in total across 88 safetensors shards. Benchmarks (NVFP4 \u002F BF16): GPQA-Diamond 89.90 \u002F 89.52, Terminal-Bench 2.1 62.78 \u002F 60.67, SWE-bench Verified 71.07 \u002F 71.10, SWE-bench Pro 51.30 \u002F 51.78, tau2-bench retail 69.30 \u002F 69.24, tau2-bench airline 75.75 \u002F 76.50, Video-MME v2 32.54 \u002F 32.85, OSWorld-Verified 52.3 \u002F 51.14, AA-Omniscience index 5.3 \u002F 4.03, GDPval-AA v2 1206 \u002F 1211. License is OpenMDW-1.1 (commercial use permitted). **Cataloged with zero on-site runs**: under the site's conservative padded VRAM accounting a 120B-class NVFP4 model needs roughly 180GB+, so 96GB-class cards do not get a fit verdict.",null,[16,18],{"url":11,"label":17},"HuggingFace",{"url":19,"label":20},"https:\u002F\u002Fhuggingface.co\u002Fnvidia\u002FNVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16","BF16",{"recordCount":22,"variantCount":22,"hardwareCount":22,"unlinkedRecordCount":22,"evidence":23},0,{"L0":22,"L1":22,"L2":22},[],{"parents":26,"children":27},[],[],[],"2026-10-24T01:42:21.448Z",{"favorited":31},false,{"items":33,"total":22,"page":34,"pageSize":35},[],1,50,{"items":37,"total":34,"page":34,"pageSize":51},[38],{"id":39,"type":40,"title":41,"excerpt":42,"author":43,"replyCount":46,"createdAt":47,"lastRepliedAt":14,"resolvedAt":14,"editedAt":14,"model":48,"hardware":49,"variant":14,"record":14},45,"SHOWCASE","Nemotron 3.5 Super-VL 120B-A12B：75GB NVFP4 权重塞进 96GB PRO 6000 的可行性讨论","NVIDIA 的另一个 Nemotron 3.5：Super-VL-120B-A12B-NVFP4，120B 总参 \u002F 12B 激活，Mamba-2 + Transformer 混合 MoE + C-RADIO 视觉编码器，1M 上下文，原",{"nickname":44,"avatarUrl":45},"ke chen","https:\u002F\u002Flh3.googleusercontent.com\u002Fa\u002FACg8ocLqEPqLilD8toFN9SqqHpAwEhXNJ0Fq8Rs8JxIKZoIZr8T_pg=s96-c",3,"2026-10-24T01:48:47.088Z",{"name":5},{"name":50},"NVIDIA RTX PRO 6000 Blackwell 96GB",20,{"items":53},[],{"items":55,"total":34,"page":34,"pageSize":46},[56],{"slug":57,"category":58,"publishedAt":59,"titleZh":60,"titleEn":61,"summaryZh":62,"summaryEn":63,"models":64,"hardwares":72},"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.",[65,68,69],{"id":66,"name":67},67,"gemma-4-E4B-it",{"id":4,"name":5},{"id":70,"name":71},97,"NVIDIA-Nemotron-3.5-Bolt-E4B-4GB",[73,76,79],{"id":74,"name":75},13,"NVIDIA RTX 5090",{"id":77,"name":78},54,"NVIDIA DGX Spark",{"id":80,"name":50},68]