NVIDIA Nemotron 3.5 Super VL

Open-weight multimodal reasoning for visual AI agents and complex workflows

Try now

Read Docs

About model

NVIDIA Nemotron 3.5 Super VL is an open-weight multimodal model designed for visual reasoning and long-running AI agents. It combines text, image, document, and video understanding with agentic reasoning and tool use in one model, preserving task-relevant visual context throughout specialized workflows. Built on a 120B parameter Mixture-of-Experts architecture with 12B active, it pairs a hybrid Mamba-Transformer design and multi-token prediction with NVFP4 KV cache quantization for efficient, concurrent long-context serving. This model can be customized through post-training on an organization's own data, tools, and policies for the highest accuracy on specialized multimodal agent tasks. Within a system of models, NVIDIA Nemotron 3.5 Super VL serves as the high-accuracy multimodal model, alongside Nemotron 3.5 Lightning for high-volume text and Nemotron 3 Ultra for the hardest text-only reasoning. Available on Together AI.

‍

Total Parameters (12B active)

120B

12B active parameters per token, balancing accuracy and efficiency for always-on agent tasks

Model Architecture

MoE

Hybrid Mamba-Transformer design supporting efficient inference and complex reasoning

Multimodal Understanding

Unified text and visual understanding for documents, images, and video

Model key capabilities

Multimodal Understanding: High-accuracy reasoning and tool use combined with document, image, and video understanding in one model
Domain-Specific Customization: Open-weight model adaptable through post-training on an organization's own data, tools, and policies
Efficient Mixture-of-Experts Architecture: 120B total parameters with 12B active per token, hybrid Mamba architecture and multi-token prediction for faster generation, with NVFP4 KV cache quantization creating headroom for more concurrent long-context sessions
Production-Ready Infrastructure: 99.9% SLA, available on Together AI serverless infrastructure

API usage

cURL
Python
Typescript

Endpoint:

nvidia/NVIDIA-Nemotron-3.5-Super-VL

curl -X POST "https://api.together.xyz/v1/chat/completions" \
  -H "Authorization: Bearer $TOGETHER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "nvidia/NVIDIA-Nemotron-3.5-Super-VL",
    "messages": [
      {
        "role": "user",
        "content": "What are some fun things to do in New York?"
      }
    ]
}'

from together import Together

client = Together()

response = client.chat.completions.create(
  model="nvidia/NVIDIA-Nemotron-3.5-Super-VL",
  messages=[
    {
      "role": "user",
      "content": "What are some fun things to do in New York?"
    }
  ]
)
print(response.choices[0].message.content)

import Together from 'together-ai';
const together = new Together();

const completion = await together.chat.completions.create({
  model: 'nvidia/NVIDIA-Nemotron-3.5-Super-VL',
  messages: [
    {
      role: 'user',
      content: 'What are some fun things to do in New York?'
     }
  ],
});

console.log(completion.choices[0].message.content);

Model card
Architecture Overview:
• Mixture-of-Experts design with 120B total parameters and 12B active per token
• Hybrid Mamba-Transformer architecture with multi-token prediction for faster generation
• C-RADIOv4 vision encoder providing native image, document, and video understanding
• NVFP4 KV cache quantization reduces cache memory footprint, supporting more concurrent long-context sessions and reducing cache eviction and recomputation in multi-turn agent workflows

Training Methodology:
• Built for high-accuracy multimodal reasoning within systems of models, supporting specialized agent workflows that require visual understanding and tool use
• Open-weight model designed for post-training on domain-specific data to improve accuracy and adapt to an organization's tasks, tools, policies, and workflows

Performance Characteristics:
• Designed for high-accuracy multimodal reasoning and complex agentic workflows, combining a 120B-parameter MoE architecture with 12B active parameters per token for efficient inference at scale
• Benchmark figures arrive with the launch materials
‍
Prompting
Together AI API Access:
• Access NVIDIA Nemotron 3.5 Super VL via Together AI APIs using the endpoint nvidia/nemotron-3.5-super-vl
• Authenticate using your Together AI API key in request headers
• Supports text, image, and video input with text output, including documents and screenshots
• Suited to multi-step tool workflows where visual context must persist across turns
• Available on Together AI serverless infrastructure
‍
Applications & use cases
Customer Service:
• Resolve multi-step requests by interpreting customer-provided documents and images
• Apply policy rules and invoke service tools within one agent session
• Keep visual evidence in context from intake through resolution

Document Intelligence:
• Interpret complex documents including OCR, tables, charts, and page layouts
• Extract needed information and generate summaries or reports
• Process document batches as always-on agent workloads

Video Understanding:
• Extract key information from video for search, content review, and incident analysis
• Support advertising and retail workflows with visual content analysis
• Combine video understanding with tool use for downstream actions

Web Research & Enterprise Productivity:
• Search across websites, connect information from multiple sources, and verify results
• Complete multi-step productivity tasks by interpreting visual information and acting on it
• Operate application workflows grounded in screenshot understanding
‍