NVIDIA Releases Nemotron 3.5 Super VL, a 120B Open Multimodal Agent Model
https://www.marktechpost.com/2026/10/15/nvidia-nemotron-3-5-super-vl-release/📌 【NVIDIA 最新釋出】Nemotron 3.5 Super VL：一個 checkpoint 扛起文字、影像與 Agent
TL;DR：NVIDIA 開源 1200 億參數多模態 MoE 模型,單一 checkpoint 同時處理文字、影像、影片與工具呼叫,免除企業維護雙模型的麻煩。
過去要做一個「永遠在線」的企業 agent,得同時維護一個文字模型和一個視覺模型,串接起來還得處理介面不一致的問題。NVIDIA 這次直接把答案寫進模型架構裡:一個開源 checkpoint,讀文字、圖片、文件、影片,然後推理、呼叫工具、寫程式碼,一次到位。
🤔 為什麼企業需要「單一模型」
Nemotron 3.5 Super VL 是一個 1200 億參數的 Mixture-of-Experts(MoE)模型,每個 token 只啟動 120 億參數。對開發者而言,這代表原本要分開部署的文字模型與視覺模型,現在可以用同一份開源權重取代。
🧩 Mamba-2、MoE、Attention 交錯的混合架構
模型骨幹交錯疊放 Mamba-2、MoE 與 attention 層,每個 MoE 層內有 512 個 routed experts。視覺輸入則靠一個 C-RADIO 視覺編碼器搭配 connector,把圖片、文件頁面與影片幀餵進語言骨幹。影片輸入限制在 2 分鐘以內的 MP4 片段,推理模式可透過 chat template 開關切換。
這個模型是從 Nemotron 3 Ultra 蒸餾而來。NVIDIA 團隊表示訓練用了約 26 兆 tokens,後訓練階段用了約 380 萬筆樣本,並在數十個環境中進行強化學習;持續預訓練階段又加入了 100 億 Indic tokens。
📊 參數少了 4.6 倍,分數反而更高
在 agent harness 評測中,Super VL 的表現超越參數量約為其 4.6 倍的 Nemotron 3 Ultra:Hermes Bench 拿下 63.44% 對比 Ultra 的 56.33%,LangChain evals 則是 87.05% 對比 84.3%。
速度方面,NVIDIA 透過 Baseten 端點在 Artificial Analysis 的 agentic 任務測試中,回報每秒超過 600 個輸出 token,圖表顯示比「Inkling Small」快上 5 倍。在 16 張 B200 GPU、每使用者每秒 50 token 的條件下,NVIDIA 宣稱能多服務 8 倍使用者,首個 token 延遲則下降 85%。這些分數皆為 NVIDIA 自行回報,不同廠商的 harness 與 benchmark 版本不一,幾分之差應謹慎看待。
💡 邊緣端也有對應版本
NVIDIA 同步推出兩個邊緣模型:Nemotron 3.5 Bolt 74B-A9B 與 Bolt E4B,同屬開源多模態 agent 模型家族。Bolt 74B-A9B 鎖定本地端 agent 應用,發布說明指出它是從 Super VL 蒸餾而來,原生 2-bit 量化後可塞進 RTX 與 DGX Spark 系統的 24GB 記憶體。NVIDIA 表示它在 PinchBench 上於同級開源模型中排名第一,在 BrowseComp 上排名第二。Bolt E4B 僅 4GB,鎖定語音助理與嵌入式 agent,預計 10 月 23 日於 GeForce RTX、Jetson 與 RTX/DGX Spark 平臺上線。
權重以 OpenMDW-1.1 授權釋出,NVIDIA 同時開源了資料集、訓練配方與 NeMo Gym 中的 RL 環境。模型可在 Hugging Face、OpenRouter、ModelScope 取得,也可用 NVIDIA NIM 部署,支援 vLLM、SGLang、TensorRT-LLM(Blackwell 與 Hopper)、llama.cpp 與 Unsloth。正式環境可搭配 Dynamo recipes 做 KV-aware routing;客製化則透過 NeMo AutoModel(SFT、LoRA)與 NeMo RL 進行。
⚠️ 分數比較需要保留懷疑
所有評測分數皆為 NVIDIA 自行回報,且來自 NVFP4 checkpoint,不同 harness 與版本差異可能影響比較結果。
🎯 實務啟示
如果你的團隊正維護一套「文字模型 + 視覺模型」的拼接架構,Super VL 提供了單一 checkpoint 的替代路線;而 Bolt 系列則讓本地端、低記憶體的 agent 部署變得可行,值得評估是否能簡化現有的多模型 pipeline。
🔗 來源
- 標題:NVIDIA Releases Nemotron 3.5 Super VL, a 120B Open Multimodal Agent Model
- 作者／機構:Asif Razzaq,MarkTechPost
- 連結:https://www.marktechpost.com/2026/10/15/nvidia-nemotron-3-5-super-vl-release/
#NVIDIA #Nemotron #OpenSource #MultimodalAI #MixtureOfExperts #AIAgent #Mamba2 #LLM #EdgeAI #VisionLanguageModel
原始資料 MarkTechPost · 收集於 2026-10-16
摘要原文
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. 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. 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. 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. Speed comes from four techniques: 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. All scores below are NVIDIA-reported, from the NVFP4 checkpoint on the model card. 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. 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. Scores are self-reported by each vendor. Harnesses and benchmark versions differ, so treat gaps of a few points with caution. Super VL is available on Hugging Face , OpenRouter , ModelScope and as an NVIDIA NIM . Supported runtimes are vLLM, SGLang and TensorRT-LLM on Blackwell and Hopper. Local paths include llama.cpp and Unsloth . For production, Dynamo recipes add KV-aware routing. On 16 B200 GPUs at 50 tokens per second per user, NVIDIA reports 8x more users served. First-token latency drops 85%. Customization runs through NeMo AutoModel for SFT and LoRA, and NeMo RL for reinforcement learning. NVIDIA cites early results: Nemotron 3.5 Bolt 74B-A9B targets local agents. NVIDIA’s launch notes describe it as distilled from Super VL. Native 2-bit quantization fits it into 24GB of memory on RTX and DGX Spark systems. NVIDIA says it ranks #1 among open models in its class on PinchBench . It ranks #2 on BrowseComp among open models. It is available now on Hugging Face with llama.cpp, Ollama, LM Studio, Unsloth and vLLM support. Bolt E4B 4GB arrives October 23. It targets voice assistants and embedded agents on GeForce RTX, Jetson and RTX or DGX Spark. Check out the technical report , model card and usage cookbook . All credit goes to the researcher of this project. Also, feel free to follow us on Twitter and don’t forget to join our 150k+ML SubReddit and Subscribe to our Newsletter . Wait! are you on telegram? now you can join us on telegram as well. The post NVIDIA Releases Nemotron 3.5 Super VL, a 120B Open Multimodal Agent Model appeared first on MarkTechPost .
由 claude-code/sonnet 自動生成