Nemotron 3.5 Super VL explained: what it changes for teams

What NVIDIA launched, and what it changes for a small team

Amazon Web Services has listed NVIDIA's Nemotron 3.5 Super VL on SageMaker JumpStart. For a small or mid-size B2B company, the category decision is whether an open vision-language model belongs in your document and screen-reading work this month, and our answer is a bounded pilot on a hosted endpoint now, with a self-run AWS deployment only once an engineer owns it.

Context: up to 1 million tokens according to Amazon Web Services (2026), in a post dated October 16, 2026 that also describes a model accepting image, video and text input and returning text, with 120 billion total parameters, 12 billion active per token, a built-in Multi-Token Prediction head for faster text generation, and two JumpStart variants: NVFP4 for G7 instances with RTX PRO 4500 Blackwell Server Edition GPUs, and BF16 reference weights on ml.p5en.48xlarge for customization.

Two details need checking first. AWS's URL carries a 2026/01 path segment while the page itself is dated October 16, 2026, so confirm availability in your own console before you set a go-live date. We found no source saying the model was renamed, merged or discontinued, and the nearest earlier multimodal model on the same platform is Nemotron 3 Nano Omni, which we compare below.

TL;DR: Nemotron 3.5 Super VL is a larger, longer-context, open-weight step up from Nano Omni, with no audio input and a different license for its weights. For most small teams the useful move is a short pilot on documents you already process, not a platform decision.

Key Takeaways

• Intelligence Index: 26.5, against 13 for Nemotron 3 Super according to Artificial Analysis (2026), which also gives 46k output tokens per task against 114k for the predecessor.

• Launch claim: up to five times faster than rivals according to IT Brief (2026). NVIDIA does not name those rivals, so treat the claim as direction, not measurement.

• Video and audio are narrower than the launch copy: NVIDIA's training documentation says video training and video inference are not verified, and the model has no audio encoder. Pilot images and text first.

• The license changed from the predecessor: the weights fall under the OpenMDW license, while AWS describes Nano Omni as covered by the NVIDIA Open Model Agreement. Have counsel read both before production use.

• Hosted prices differ by provider: one listed rate is not the same as the median across providers, so check the provider you route to before you budget.

• Our recommendation: pilot on a hosted endpoint this month, and name an engineering owner before any AWS endpoint goes live.

How we evaluated these options

We scored the options on six factors that a small B2B team actually decides on, and weighted each one by how often it would change a go or no-go call. The weights are our judgment rather than a published standard, so change them if your team weighs cost or compliance differently.

We compare two ways to run Super VL, on SageMaker JumpStart and as a hosted endpoint, against Nemotron 3 Nano Omni, the earlier multimodal model on JumpStart. We did not open other vendors' pages, so we do not rank other models here.

Signal vs Speculation

What the sources confirm. The parameter counts, the context window, the JumpStart variants, the license named for the weights, and the absence of an audio encoder all appear on pages cited in this article. Listed prices and throughput depend on the provider and the day, which is why we show both a listed rate and a median.

What remains vendor claim. NVIDIA says Super VL delivers up to five times the throughput of comparable open models. The coverage we read does not name the comparison models or publish the scores behind that statement. AWS lists video question answering as a use case, while NVIDIA's documentation says video inference is not verified, and the two should be read together.

Our reading, not a fact. For a small team, the likely main effect is a cheaper first test of a document or screen workflow, because listed token prices are low enough that human review becomes the bigger cost line. We expect the license and the deployment path to decide more go or no-go calls than the benchmark does. Public data cannot settle either point, so check both against your own documents and your counsel. Artificial Analysis also rounds its score differently on its model page than in its launch article, so quote the score with its date and source.

Feature matrix: Super VL against its predecessor

"Not in the sources we opened" means the figure is absent from the pages we read, not that it is zero. The sources for these figures are given in the text of this article.

Pricing and total cost

Pricing checked October 17, 2026.

Hosted list price: $0.10 input, $0.40 output per million tokens according to OpenRouter (2026), where NVIDIA is the only provider the page names. The median across providers is higher, at $0.23 input and $0.57 output per million tokens, according to Artificial Analysis (2026), and that same page gives a median output speed of 388.0 tokens per second.

On SageMaker, the AWS pricing page we opened says there is no additional charge for JumpStart models and that you pay for the instance type you choose. It does not list per-hour rates for ml.p5en.48xlarge or the G7 family, so check the AWS Pricing Calculator for those rates before you commit a budget.

The table below is an illustrative calculation, not a measured bill. It uses the token volumes from the worked example later in this article.

Per-option profiles: best fit, limits and evidence

Super VL on SageMaker JumpStart

Best fit: a team that already runs on AWS, has an engineer who can deploy and monitor an endpoint, and wants the model to run inside its own cloud account. The limits are cost and effort. You pay for instance hours, and the BF16 reference weights are the customization path, not the cheap inference path. BF16 checkpoint: about 248.9 GB according to NVIDIA (2026), and the same documentation says its LoRA recipe was verified on 16 H100 GPUs. It also says video training and video inference are not verified, and that native Hugging Face video forward currently fails. Deploy from the JumpStart catalog in the SageMaker console or through the SageMaker Python SDK, and check G7 or p5en quota in your account before you set a launch date.

Super VL as a hosted endpoint

Best fit: a pilot team with no GPU capacity and an engineer who can call a chat completions API. The limits are residency and variance. The OpenRouter page we opened names NVIDIA as the only provider and says nothing about EU in-region routing, so check where your documents would be processed before you send customer files. Provider prices and speeds vary, and the median output speed is an average across providers, not a guarantee for your prompts. For another model choice framed around cost and speed, see a separate comparison of cheaper alternatives to GPT-5.6 Sol.

Nemotron 3 Nano Omni, the previous multimodal model

Best fit: a team that needs audio processed in the same model, such as recorded calls, meetings or voice notes, and does not need the longest context window. Context: 131K tokens according to Amazon Web Services (2026), which also describes 30 billion total and 3 billion active parameters, video, audio, image and text input, FP8 precision on JumpStart with ml.p4d.24xlarge and ml.p5.48xlarge as example instances you may need quota for, a post dated 28 April 2026, and a license it describes as the NVIDIA Open Model Agreement, which covers commercial use. Its limits are a shorter context and a license that is a separate agreement from Super VL's, so the license review is a separate task.

Document intake is where the choice shows up

Invoice intake is where this model choice becomes concrete, so the design below is a proposed, configurable flow rather than something running today. US Tech Automations could configure a flow that starts when a vendor invoice PDF lands in a shared accounts-payable mailbox. The flow renders each page as an image, sends it to the model with a JSON schema covering vendor name, invoice number, due date, currency, total and line items, and receives a structured record back, since the model supports structured JSON output and tool calling as the launch material describes.

Before anything is written, the flow compares the total with the matching purchase order and sends any mismatch, any new vendor and any record with a blank required field to a human reviewer. On approval, it creates a draft bill in QuickBooks Online through that product's API. The prerequisites are a model endpoint with an access key, a schema agreed with your accounts-payable lead, an OAuth connection to QuickBooks Online, and a set of past invoices your team has already labeled for checking. The review points are the mismatch queue, a monthly sample of approved records, and a sign-off on the field mapping by whoever owns the books.

Here is the arithmetic for one illustrative scenario, not a measured result. Assume 2,000 invoice pages a month, each sent as one image with about 1,500 input tokens and returning about 600 output tokens. That is 3 million input tokens and 1.2 million output tokens. At the listed OpenRouter rate of $0.10 and $0.40 per million tokens, the model charges come to $0.78 a month. If 10 percent of pages, or 200, need four minutes of human review, that is 800 minutes, or about 13.3 hours a month, and review time, not the token charge, sets the real cost. On the request side, NVIDIA's API reference for the text‑input Nemotron 3 Super documents enable_thinking under chat_template_kwargs for controlling reasoning, so confirm that the Super VL endpoint accepts the same field before you build.

Build it yourself, or use a workflow layer?

Your real alternative is often a flow in Zapier, Make or n8n, or a script your engineers write. Those tools can keep run histories, retries, error branches and audit evidence when they are configured that way. The buyer still designs and owns idempotency, so a retried invoice does not create two bills, along with escalation paths, access controls, and upkeep when the model endpoint or the QuickBooks Online API changes. A configured US Tech Automations design could differ mainly by packaging the same steps, extraction, purchase-order matching, the review queue and the approval record, as one flow with the prerequisites and review points listed above. Whether that beats your own build depends on how much engineering time you have, and we would not assume it does.

When NOT to use US Tech Automations

A workflow layer is the wrong choice when you process a few dozen invoices a month and a person can read them faster than a flow can be set up, when your accounting system's built-in capture already reads your invoice format, or when the job is a one-off prompt that someone runs by hand in an AI assistant your team already pays for. In those cases the model choice matters less than the habit of checking each output, and a simpler tool wins on setup time and cost.

Who this is for

This fits a small B2B software or service company with a document-heavy or screen-heavy process, such as invoice intake, contract review, or support tickets that include screenshots, plus at least one engineer who can call an API or run an endpoint. It is not the right first step for a team that wants a finished app with no configuration.

Red flags: no one on the team can own an endpoint, its quota and its access keys; your documents must stay in a region that the hosted provider does not serve; or your main need is audio, which this model does not take as input.

Common mistakes

• Treating the long context window as proof that very long documents will work. NVIDIA's documentation records no verified result for long-context supervised fine-tuning.

• Counting video as a production feature because the launch copy lists it. NVIDIA's documentation says video inference is not verified.

• Picking the BF16 variant for everyday inference. AWS positions BF16 for customization and NVFP4 for optimized inference.

• Budgeting from token prices alone, which hides the review hours shown in the worked example. For a broader view of costing a tool stack, see how SaaS teams save on workspace tools.

• Reading the Intelligence Index as a score for your own forms. It measures general tasks, not your documents.

Glossary

• Active parameters: the subset of a Mixture-of-Experts model's parameters used to process each token.

• Context window: the amount of input the model can take in a single request.

• Multi-Token Prediction (MTP) head: a built-in component used for speculative decoding, which speeds up text generation.

• NVFP4 and BF16: NVFP4 is NVIDIA's 4-bit floating-point format for optimized inference, and BF16 is a 16-bit format used here as the customization reference.

• JumpStart: AWS's model catalog for deploying models into your own account, with no added charge for the model itself, though you pay for instance hours.

• OpenMDW license: the license NVIDIA's catalog names for the Super VL weights. Read it before production use.

• Intelligence Index: Artificial Analysis's composite score across a set of tasks, useful for comparing models and not a substitute for testing on your own documents.

Frequently asked questions

Can a company use Nemotron 3.5 Super VL commercially?

Possibly, but the sources do not settle it: NVIDIA's catalog page says the weights are governed by the OpenMDW License Agreement, version 1.1, and does not spell out commercial-use terms, so have counsel read the license before production use.

Does Nemotron 3.5 Super VL process video and audio?

It takes video as an input according to the launch material, but NVIDIA's documentation says video training and video inference are not verified, and the model has no audio encoder, so treat video as a pilot question and leave audio out of scope.

How is it different from Nemotron 3 Nano Omni?

It is larger, has a longer context window, drops audio input, and uses a different license agreement for its weights, which is why the license review is a separate task.

Should a small team host it on AWS or call a hosted endpoint?

Start with a hosted endpoint for a pilot, because a SageMaker deployment brings instance-hour billing, quota work and an owner for the endpoint. Move to AWS when the pilot shows a volume or data-control reason that justifies that work.

What does it cost to run each month?

In the worked example, the listed token charges are small, and review time is the larger cost to plan for, so budget for people before you budget for tokens.

Are NVIDIA's speed and accuracy claims proven?

Not by the sources we read: NVIDIA's claims are company statements, and the independent Artificial Analysis index is the stronger signal, though it measures general tasks rather than your forms.

What to do this month

Run a bounded pilot this month. Pick a sample of invoices or screens your team already handles, measure review time with and without the model, and choose the hosting path only after you have that number. Read the OpenMDW license with counsel before any customer documents pass through the model. If you want a concrete picture of how this kind of intake flow would be configured, see how US Tech Automations configures this. For a parallel model decision aimed at SaaS teams, see GPT-6 Astra vs Claude Fable 5.1 for SaaS.