• 運算
  • 客戶
  • 價格
登入
More Blog Posts
XDiscordLinkedInYouTube

產品

  • GPU
  • MaaS
  • Studio

開發者

  • 模型總覽
  • 技術文件
  • 詞彙表

公司

  • 關於我們
  • 部落格
  • 活動
  • 合作夥伴
  • 新創計劃
  • 職涯
  • 大使計畫
  • 使命與願景

熱門模型

    掌握 AI 最新動態

    提交即表示您瞭解我們會收集並使用您提交的資訊,其中可能包含個人資訊。

    XDiscordLinkedInYouTube

    Copyright ©2026 All rights reserved.

    隱私政策使用條款法律文件
    More Blog Posts
    Research

    Nemotron 3.5 Super VL Watched Six Security Feeds All Day

    NVIDIA’s open Nemotron 3.5 Super VL monitored six security cameras on GMI Cloud for about 20 cents an hour. Specs, benchmarks, and how it compares to Qwen and DeepSeek.

    2026年10月23日

    A 120-billion-parameter open model watched six security feeds, raised two alerts, both correct, and billed at a rate of about 20 cents an hour of footage. I built the agent, measured the lag and the bill, then put the same clip through the two open Flash models people would reach for instead. The video shows it running. This post covers what the numbers say.

    The model

    Nemotron 3.5 Super VL is the vision member of NVIDIA's Nemotron 3.5 Super line. It keeps the Nemotron Super hybrid decoder and puts a RADIO v2.5-H vision encoder in front of it. NVIDIA's documentation lists 121B total parameters with 512 routed experts and 22 active per token, which is where the 12B-active figure in the name comes from.

    Nemotron 3.5 Super VL

    Total / active parameters

    121B / 12B

    Decoder

    88 layers: 40 Mamba-2, 40 mixture-of-experts, 8 attention

    Experts

    512 routed, 22 active per token, plus 1 shared

    Vision

    RADIO v2.5-H, 512×512 tiles, 256 tokens per tile, separate video embedder

    Audio

    none

    Thinking

    off, low effort, full, set through the chat template

    License

    NVIDIA Nemotron Open Model License, commercial use allowed

    Price on GMI, per 1M tokens

    $0.30 in / $0.90 out, list estimate

    Two details in that table decide the rest of this post. Only 8 of the 88 layers are attention, so cost grows slowly as the input grows; the family's one-million-token context claim rests on that. And each of the three thinking settings returns a reasoning token count in the usage block, so you can price every setting on your own prompts instead of guessing.

    The deployment I used was a private GMI endpoint ahead of the public release, and it accepted about 32K tokens of context. Nothing below depends on more than that.

    Three benchmarks, and the caveat

    NVIDIA's published benchmark table covers the text backbone, Nemotron 3 Super; the VL card was still gated when I ran these tests. The two models you would pick instead on GMI are DeepSeek-V4-Flash and Qwen3.8-Flash: open weights, flash prices, both newer.

    Benchmark

    Nemotron 3.5 Super VL (backbone)

    DeepSeek-V4-Flash

    Qwen3.8-Flash

    GPQA Diamond, thinking on

    79.2

    87.4

    91.7

    SWE-bench Multilingual

    45.8

    70.2

    81.0

    LiveCodeBench

    81.2 (v5)

    88.4

    91.9 (v6)

    RULER at 1M tokens

    91.75

    not reported

    not reported

    List price, in / out per 1M tokens

    $0.30 / $0.90

    $0.14 / $0.28

    $0.16 / $0.47

    The two newer models win every text row, and Nemotron is the most expensive of the three. The harnesses differ (Nemotron's SWE-bench run is OpenHands, Qwen's is mini-SWE-agent, DeepSeek does not say, and DeepSeek's rows are its High thinking mode), so I would treat gaps under five points as noise. These gaps are not under five points.

    Where the backbone stands out is long input. It scores 91.75 on RULER at one million tokens against 22.30 for GPT-OSS-120B, and NVIDIA reports up to 2.2 times the throughput of GPT-OSS-120B at 8K tokens in and 64K out, on its own hardware. Neither number appears in a leaderboard screenshot. Both showed up in the agent below.

    What I built

    Lookout is one browser page with six camera feeds. Four show ordinary activity: a corridor, a control room, a plant hallway, a loading dock. Two carry staged incidents. Every few seconds each camera records its last clip in the browser and sends it to the model as a video_url part, with a rulebook written in plain English: staff at desks and people walking corridors are normal; a covered face indoors, forced entry, opening a rack without a badge, or covering the camera is an alert.

    The model answers with one JSON object per clip: type, severity, who, what they are doing, where in the frame, which rule fired, the indicators it saw, and a recommended action. The box on the feed is drawn from the coordinates the model returns. There is no separate detector.

    Measured on the GMI endpoint, October 8, 2026

    Cameras on the wall

    6

    Clip length

    6 s

    Prompt tokens per second of 512px video

    about 90

    Lag behind live

    about 2 s per clip

    Alerts raised

    2, both on the staged cameras

    False alerts on the four ordinary feeds

    0 after one rule change

    Cost per hour of footage

    $0.17 to $0.26 across runs, at the $0.30 / $0.90 estimate

    Thinking

    off

    Two things from those runs that the video does not show. The one false alert I hit came from the corridor feed, where a backlit silhouette was read as a covered face. One added line in the rules, that a dark silhouette or a backlit figure is not a covered face, fixed it on the next run. That is the argument for writing rules in prose instead of training a detector: the fix took a sentence, not a dataset. The second is that the model will describe an image you did not send. Asked about a picture with no picture attached, it invented one. Attach the frame or do not ask.

    Two limits to design around: one image or one clip per message, because two image parts in one request returned a server error; and stream anything long, because the edge in front of the endpoint closes a connection that stays silent for two minutes.

    Against two open Flash rivals

    I tried to run Lookout on the other two. DeepSeek-V4-Flash returns HTTP 400 on GMI for any image or video part, so it is out for this job whatever its benchmark lead. Qwen3.8-Flash reads video and described the same six-second clip correctly, but it refused the browser's WebM recordings, so it could not run on the live wall; I had to transcode a clip to test it at all.

    Same 6-second clip

    Nemotron 3.5 Super VL

    Qwen3.8-Flash

    DeepSeek-V4-Flash

    Accepts video on GMI

    yes

    yes, not WebM

    no, HTTP 400

    Prompt tokens for the clip

    491

    944

    n/a

    Input price per 1M tokens

    $0.30

    $0.16

    $0.14

    Input cost for the clip

    $0.00015

    $0.00015

    n/a

    The token row is the one that matters for a camera wall. Nemotron's encoder spends about half the tokens Qwen's does on the same footage, and at list prices that makes the two a wash per clip, because Nemotron's rate is about double. Where they separate is capacity. Half the tokens per clip means twice the cameras per unit of context and throughput, and a clip the model takes straight from the browser means one less transcoding step in the loop.

    Where I would route it

    • A camera, a screen or a document at the front, JSON at the back. Lookout-style loops, UI screenshots, scanned forms. This is the job one rival cannot do and the other cannot do from a browser clip.

    • Volume agent work. Ticket queues, batch extraction, tool loops. In a separate test on the same endpoint, 24 agents resolved 24 IT tickets at once at 348 tokens per second across the swarm, with thinking on low effort.

    • Anywhere you want the weights. NVFP4 and FP8 checkpoints exist, the license allows commercial use, and the BF16 checkpoint fits on two B200s. You can leave GMI with it, which is part of why I am comfortable renting it there.

    • Not for the leaderboard. On reasoning and multi-file coding, DeepSeek-V4-Flash and Qwen3.8-Flash are ahead by 7 to 35 points at lower prices.

    • Thinking off by default. For triage, extraction and tool loops it was faster and cheaper, and the usage block tells you what each setting cost. Try low effort before full.

    Try it on GMI Cloud

    One base URL, one key, and the OpenAI request shape. The only additions are a video_url part in the message and the thinking toggle in the chat template.

    curl https://api.gmi-serving.com/v1/chat/completions \
      -H "Authorization: Bearer $GMI_API_KEY" \
      -H "Content-Type: application/json" \
      -d '{
        "model": "nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16",
        "messages": [{"role": "user", "content": [
          {"type": "video_url", "video_url": {"url": "data:video/webm;base64,..."}},
          {"type": "text", "text": "What happens in this clip? Reply as JSON with type, severity, who, what, where."}
        ]}],
        "response_format": {"type": "json_object"},
        "chat_template_kwargs": {"enable_thinking": false},
        "stream": true
      }'

    Three notes from the runs. One image or one clip per message. For a short reasoning trace, set enable_thinking to true and add "low_effort": true in the same block. And keep stream on for anything long. To rerun the text comparison, swap the model field for deepseek-ai/DeepSeek-V4-Flash or Qwen/Qwen3.8-Flash; only Qwen will accept the video part.

    The question to answer

    Do not ask whether Nemotron beats the newer Flash models on a leaderboard. It does not. Ask which of your loops has a camera, a screen or a document at the front and a JSON object at the back, and what that loop costs per hour today. Build it with the request above, watch the usage block, and put the hourly number next to whatever you pay now.

    Ready to build? Get Started

    Grace Deng

    DevRel @ GMI Cloud

    Build AI Without Limits

    GMI Cloud helps you architect, deploy, optimize, and scale your AI strategies

    Ready to build?

    Explore powerful AI models and launch your project in just a few clicks.

    Get Started