One model for the reading and the doing

Most enterprise agent work is not a chat problem. A credit review, a claims check or a due-diligence pass means reading a PDF, pulling a figure from a table, checking a chart against a webpage, and only then calling a tool or writing an answer. Teams have usually split that job between a vision-language model that reads and a language model that reasons, passing context between them and losing some of it on the way.

Super VL puts both halves in one system. Its vision training is aimed at the material agents actually meet in offices: optical character recognition, tables, charts, multipage documents and structured information on webpages. NVIDIA says that is the point of the release. A single model can follow the document from evidence to decision, and an organisation can post-train it on its own files, policies and procedures, then run it where its data-residency rules require.

That pitch is concrete enough to test. Either the model extracts the right figure from your agreements and follows your review rules, or it does not. No leaderboard settles that for a buyer.

The 120B number, read properly

Super VL stores 120 billion parameters and uses about 12 billion of them for each token. That mixture-of-experts design is legitimate, and it is also why neither figure can be read the way a dense model's would be. For memory it is not a 12B model: memory has to hold the whole 120B, which is why the model card's deployment list starts at a DGX Spark, an H200, two H100s or a GB200 for the NVFP4 checkpoint, rather than at an ordinary workstation. For per-token compute it is not a dense 120B model either: each token engages only the active subset, so the arithmetic per token sits closer to the 12B figure.

The published recipes are not interchangeable, either. The model card's deployment table lists the DGX Spark's vLLM configuration as “Text, image” — no video input — while the vLLM recipes for the H200, the GB200 and two H100s serve “Text, image, and video”, and the SGLang recipe is text and image only. The architecture accepts video; the desktop recipe does not serve it. The 1M-token context comes with the same kind of condition: it is a documented maximum, with deployment settings listed on the card, and the card notes that individual recipes may use a smaller context window to fit the available GPU memory and concurrency. It is not a promise of equal performance, speed or cost at every length.

The licence is the other term worth reading before the benchmarks. The model card places the weights under OpenMDW-1.1, and NVIDIA says datasets and training recipes are open as well. For an enterprise team, that combination matters more than a small benchmark lead: it can be adapted, audited in-house and kept on premises. It is still a specific licence with its own terms, not a synonym for unrestricted use.

Speed is a chart condition, not a workflow promise

Two different speed measurements sit behind NVIDIA's launch framing, and they should not be merged. The first is NVIDIA's own claim: up to five times the output throughput of comparable open models in its size category, measured under NVIDIA's conditions. The second is a separate chart. VentureBeat's reading of the launch material describes an Artificial Analysis chart plotting index score against output speed at 100,000 input tokens; Super VL sits near 625 output tokens per second at an index score of about 26, measured on four Blackwell B200 GPUs, and other open models in the same chart score higher on the composite index.

Both numbers measure how fast tokens come out under their stated conditions. Neither measures task time. An agent that reads slowly, calls tools, checks its own work and corrects errors can spend most of a job doing things the tokens-per-second figure does not measure.

The specialised wins need the same care. NVIDIA reports improvements of between 9 and 44 percent after organisations separately customised Super VL for five specialised tasks, including document intelligence, security work, receipt extraction and medical claims. In the two comparisons VentureBeat spells out, a customised version scores well above a named Claude or GPT model on that one task. Those are different custom models, trained for different jobs. They show what post-training can buy on a narrow workflow. They do not show the base model beating frontier models in general.

What to do with it

The named adopters point at the intended use. Scale AI, AfterQuery, AI21, Mercor and Ramp are among the organisations NVIDIA says are adapting the model for finance, security, coding and document work, while Atlassian, CrowdStrike, Perplexity and Salesforce are among the platform companies integrating it. Treat those names as a map of where to look for early evidence, not as evidence itself — every result quoted so far comes from NVIDIA or from partners featured in its launch.

Run the narrow test instead. Take one document-heavy workflow your team already scores by hand, against answers you know are right, and run the released weights on it unchanged. Post-train only if the base result is close. Price the full programme only after that test: the hardware underneath it, your own evaluation and customisation time, and — only if you take the production NIM route — the AI Enterprise licence VentureBeat reports at $4,500 per GPU per year, as checked on 15 October 2026. Running the open weights yourself does not by itself require that subscription.

Sources

  1. [1] NVIDIA Blog, “Open From Data Center to Desktop: NVIDIA Expands Nemotron Family With Highest Efficiency Super VL and Bolt Models” (primary source; published 15 October 2026, 15:00 UTC — release, availability, training totals, scores, adopters, NIM and Dynamo figures)Read source
  2. [2] NVIDIA Developer Blog, “NVIDIA Nemotron 3.5 Super VL Powers Specialized Agents with Multimodal Reasoning” (primary technical source; published 15 October 2026 — architecture, CharXiv, LiLiCorr, NVFP4, context reuse and deployment detail)Read source
  3. [3] Hugging Face model card, nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4 (primary model card — 120B/12B active, Mamba-2–Transformer hybrid with C-RADIO vision encoder, up to 1M context, OpenMDW-1.1 licence, deployment hardware and per-recipe configurations)Read source
  4. [4] VentureBeat, “NVIDIA's new, open Nemotron 3.5 Super VL with post-training beats Claude Opus 5, GPT-6.1 Sol on specialized task benchmarks” (15 October 2026 — independent reading of the speed chart and post-training chart, composite index context, and the $4,500 per-GPU AI Enterprise figure)Read source
  5. [5] Unite.AI, “NVIDIA Releases Open Nemotron 3.5 Super VL and Bolt Edge Models” (15 October 2026 — corroborating release summary)Read source