NVIDIA
NVIDIA’s vision-language reasoner for text, images and visual documents. The app uses paid OpenRouter hosting; context limits vary by serving host.
At a glance
API model ID: nvidia/nemotron-3.5-super-vl
Capabilities & boundaries
Tool support requires the appropriate API integration; a supported tool is not automatically active in every chat.
NVIDIA documents configurable reasoning through its deployment chat template, including off and two thinking modes. OpenRouter advertises reasoning support, but a self-hosted enable_thinking flag is not automatically the same as a hosted request parameter. Use the selected endpoint’s supported reasoning schema and compare evidence quality, latency and token use on a representative visual task. There is no universal hosted effort scale or default established by the model card alone.
Put it to work
Original examples from EZ Ai Assist. Adapt these to your task and the features available in your workspace.
Workflow 01
Tie each conclusion to visible labels, scales and values.
Read the chart image and the written claim below. Identify the title, axes, units and time range before assessing the claim. Quote the visible values that support or contradict it, describe any scale or missing baseline that changes the interpretation, and mark unreadable labels as uncertain. Return a short evidence table and a corrected statement supported only by the supplied chart.
Workflow 02
Preserve source rows and flag uncertain or inconsistent values.
Extract the line items from this page screenshot into a table with description, quantity, unit price and stated total. Preserve currency and decimal precision exactly. Use unknown for unreadable cells, show the source row for every item, and flag totals that do not match the visible arithmetic. Do not guess hidden rows or claim that your output was validated by an enforced JSON schema.
Workflow 03
Separate visible design evidence from behavior needing a browser test.
Compare this interface screenshot with the accessibility requirements pasted below. Identify visible problems with labeling, hierarchy, touch-target spacing and contrast, while marking measurements that require source styles or an interactive browser. Cite a specific region for each observation and suggest a focused acceptance test. Do not claim that keyboard navigation or screen-reader behavior was tested from the screenshot.
Developer reference
These are published provider API reference prices, not EZ Ai Assist subscription prices or verified customer charges.
View EZ Ai Assist plans| Token type | Price |
|---|---|
| Input | $0.15 |
| Cached input | $0.02 |
| Output | $0.50 |
Common questions
Super VL is the separately identified vision-language model nvidia/nemotron-3.5-super-vl. Its model card includes visual input and a vision encoder. The existing Nemotron 3 Super guide describes a different text route; adding VL does not rename or replace that earlier model.
NVIDIA develops the model. OpenRouter exposes the paid nvidia/nemotron-3.5-super-vl route through Chat Completions and lists serving endpoints with their own limits, prices and supported parameters. The developer, gateway and serving host are distinct roles.
Only after checking the serving endpoint. The model supports up to 1M in suitable deployments, but the reviewed OpenRouter endpoints include both 262,144 and 1,048,576-token context limits. Input, visual tokens, history, reasoning and output must fit the applicable shared budget.
NVIDIA documents visual-document understanding through page images and video understanding through sampled frames. Use readable pages, retain page or timestamp references and check the serving endpoint’s input format. Visual understanding returns text rather than a new document or video file.
Use the reasoning parameters supported by the exact hosted endpoint. NVIDIA’s self-hosted chat-template controls are a different interface and do not establish a universal OpenRouter effort scale. Compare a bounded visual task before increasing reasoning expenditure.
The reviewed Together endpoint lists 235,929 maximum completion tokens within a 262,144-token context; CoreWeave lists 943,718 within a 1,048,576-token context. Those are hosted reference ceilings. The request’s input, visual tokens, history and reasoning leave less room for the final answer.
No. Small text, ambiguous scales and dense layouts can lead to confident mistakes. Ask for evidence tied to visible labels and source rows, allow unknown values and check important arithmetic against the source.
No. The table is a dated Together-on-OpenRouter API reference. Other serving endpoints publish different prices. EZ Ai Assist subscription allowances and usage charges are separate from the provider’s published token rates.
Check the source
Specifications and API prices checked on . Example prompts and workflow advice are editorial guidance from EZ Ai Assist.
Same provider
NVIDIA
NVIDIA’s flagship Nemotron reasoner for complex text, code and multi-step analysis, using a paid 256K OpenRouter route.
NVIDIA
Balanced Nemotron workhorse for text, coding and tool-assisted workflows, using a paid 256K OpenRouter route.
NVIDIA
Efficiency-focused Nemotron model for low-cost text workflows and rapid iteration, using a paid 256K OpenRouter route.
NVIDIA
Compact Nemotron model for short replies, extraction and bounded reasoning, using a paid 256K OpenRouter route.