# Build Visual AI Agents With NVIDIA Nemotron 3.5 Super VL

## Executive summary

NVIDIA Nemotron 3.5 Super VL is an open-source, multimodal LLM designed for efficient agent deployment, featuring 120 billion total parameters and 12 billion active parameters. The model combines multimodal reasoning, tool use, and code generation, allowing it to process text, images, charts, documents, and video. A demonstration showcases an AI agent using Nemotron 3.5 Super VL to investigate a synthetic billing issue, successfully identifying a 30% cost increase and generating a Python-based proposal to optimize ticket routing, projecting $100,000 in monthly savings.

## Key takeaways

- Model Architecture and Efficiency: Nemotron 3.5 Super VL is an Ultra-class model, distilling the capabilities of the Nemotron 3 family. It maintains high performance while fitting on a single NVIDIA DGX Spark, making it efficient for agent deployment. (00:00)
- Multimodal Capabilities: The 'VL' (Vision-Language) capability allows the model to parse complex documents, read dashboard screenshots, and interpret scanned pricing sheets. (00:31, 01:54)
- Automated Billing Investigation: The agent was prompted to investigate a 30% increase in AI billing costs. It read the dashboard metrics (Spend up 30%, tickets up 3%), analyzed the price list, and pinpointed the cause: a switch to a thinking model using over 100 times more thinking tokens daily. (01:09, 02:01)
- Code Generation and Proposal: The agent autonomously wrote Python code to apply a defined quality bar, comparing routing options, and generated a formal proposal that projected staying within quality standards while saving approximately $100,000 monthly. (02:09, 02:16)

## Technical details

- Model Specifications: Nemotron 3.5 Super VL is an open-source model with 120 billion total parameters and 12 billion active parameters. It is designed for fast, efficient agent execution. (00:00)
- Multimodal Reasoning Workflow: The agent workflow involves: 1) Reading dashboard metrics from a screenshot (e.g., identifying a 30% spend increase). 2) Parsing structured data from a scanned image (price list). 3) Identifying the root cause (e.g., excessive thinking token use). 4) Generating and executing Python code to optimize routing based on a quality bar. (01:46, 02:09)
- Deployment Environment: The model is capable of running on a single NVIDIA DGX Spark. (00:00)

## Practical implications

- Automating complex financial or operational investigations by combining visual data parsing (screenshots, scans) with logical reasoning.
- Implementing cost-saving agents that can analyze usage patterns and propose code-based optimizations (e.g., routing changes) while maintaining defined quality metrics.
- Integrating multimodal LLMs into CI/CD pipelines for automated report generation and anomaly detection across diverse data types (text, image, chart).

## Topics

Multimodal AI, Large Language Models (LLMs), Agentic Systems, Cost Optimization, Python Code Generation, NVIDIA Nemotron 3.5 Super VL Tech Blog, Nemotron 3.5 Super VL Download, Super VL Cookbook

Source: https://www.youtube.com/watch?v=f_rkx4coYOQ