Introduction
Nemotron 3.5 Super VL is NVIDIA's vision-language model, released on 15 October 2026 with 120B total and 12B active parameters. Here it is one chat endpoint that takes text, images and video, reads up to 262,144 tokens of input, and returns up to 262,144 tokens per reply with tools, streaming and optional reasoning.
| Field | Value |
|---|---|
| Input | Text |
| Output | Text |
| Input budget | 262,144 tokens |
| max_tokens ceiling | 262,144 |
| Inputs | text / images / video |
Key features & highlights
Built to read what is on the screen
Nemotron 3.5 Super VL pairs a language model with a vision encoder, so a chart, a form, a table or a GUI screenshot is read directly rather than through a separate OCR step. Ask for labels, values, trends and what cannot be established, and say which image each question is about.
Video as part of the conversation
A user message can carry video_url parts next to text and images. NVIDIA describes video input as sampled frames, so it suits what-happened and what-changed questions better than frame-exact timing.
Reasoning you can switch off
reasoning_effort takes none, low, medium or high and defaults to high. Use none or low for quick extraction and captions, and high when the task needs several steps. Thinking is part of the reply, so set max_tokens with room to spare.
Open weights, hosted behind one API
The model is published by NVIDIA under the OpenMDW license and built for agent work, so tool calling is first-class: up to 128 function definitions per request, parallel calls on by default, results returned as tool messages.
Request parameters
| Parameter | Required | Description |
|---|---|---|
messages | Yes | Send Nemotron 3.5 Super VL an ordered conversation containing the source material for chart and visual-document analysis; keep tool results paired with |
max_tokens | No | For Nemotron 3.5 Super VL, |
reasoning_effort | No | Nemotron 3.5 Super VL accepts none, low, medium, high, with high as the default; compare effort settings on chart and visual-document analysis while monitoring output-token usage. |
stop | No | Stop Nemotron 3.5 Super VL with one sequence or up to four sequences; the matched marker is omitted from the final text. at least 1 characters · at least 1 items · at most 4 items |
temperature | No | Use temperature to adjust sampling in Nemotron 3.5 Super VL; keep its default while evaluating chart and visual-document analysis unless you need a different answer style. |
top_p | No |
|
tools | No | Give Nemotron 3.5 Super VL up to 128 OpenAI-format function definitions for chart and visual-document analysis; your application checks arguments and performs the proposed opera… at most 128 items |
parallel_tool_calls | No | In a Nemotron 3.5 Super VL turn, this switch permits proposals for several independent functions; the application still controls their execution. Default: |
tool_choice | No | For Nemotron 3.5 Super VL, |
response_format | No | Request text or structured JSON from Nemotron 3.5 Super VL for chart and visual-document analysis, then check the JSON against the fields your application expects. |
Pricing
Input $0.20 / Output $0.50 per 1 million tokens
Use cases
Pull structured data out of documents
Send scanned pages, invoices or slides as images and ask for fields as JSON with a response_format or a tool schema. Tell the model to mark values it cannot read instead of guessing, then validate the result.
Describe and audit charts and dashboards
Attach a chart or a dashboard screenshot and ask for the labels, the trend and any number that cannot be established from the picture. Follow up with a second image to compare two periods.
Understand a screen recording
Send a short product walkthrough as video and ask for the steps shown, the error that appeared and where it first happened. For exact timestamps, extract frames yourself and send them as images.
Power a vision agent
Show the model a screenshot, let it request a tool call such as click or search, run it in your code and send back the next screenshot. Cap the number of loop steps and log every call.
Prompt guide & pro tips
Say which picture a question is about
Label images and clips in your text, for example the first chart is Q1 and the second is Q2. Keep your instructions apart from any text visible inside the image.
Pick a reasoning setting to match the job
The default is high. Drop to low or none for extraction, captions and classification, and keep high for multi-step analysis and tool loops.
Set sampling for the task
NVIDIA recommends temperature 1.0 and
top_p0.95 for text and temperature 0.6 withtop_p0.95 for vision work. Both fields are accepted here; change one at a time.Plan for the image budget
Large images and long videos use many input tokens. Resize screenshots to the size that keeps the text legible, and send sampled frames for long clips.
Nemotron 3.5 Super VL is NVIDIA's vision-language model, released on 15 October 2026 with 120B total and 12B active parameters. Here it is one chat endpoint that takes text, images and video, reads up to 262,144 tokens of input, and returns up to 262,144 tokens per reply with tools, streaming and optional reasoning.
For a chart-reading task with Nemotron 3.5 Super VL, attach the chart and request its labels, the trend and any values that cannot be read reliably. Keep observations separate from estimates. The vision-language endpoint accepts image and video parts, but not audio; it produces text rather than a revised image. Ask for a tool call only when an external calculation or lookup is needed, and verify its arguments before execution.
Content decisions come from the selected model. Nemotron 3.5 Super VL: NSFW behavior can vary, and no unrestricted-content guarantee is made.
API quick start
curl --fail-with-body -sN https://api.spicyapi.ai/v1/chat/completions \
-H "Authorization: Bearer $SPICY_API_KEY" \
-H "Content-Type: application/json" \
-H "Idempotency-Key: $(uuidgen)" \
-d '{
"model": "nvidia/nemotron-3.5-super-vl/chat",
"messages": [
{
"role": "user",
"content": "Summarise the following release notes into three bullet points."
}
],
"stream": true
}' \
| jq --unbuffered -Rrj 'select(startswith("data: ")) | .[6:] | select(. != "[DONE]") | fromjson | if .error then "\n[error] \(.error.message)\n" else (.choices[0].delta.content? // empty) end'
See the API reference for authentication, the complete schema, polling, and error handling.