Nvidia releases Nemotron 3.5 Super VL for enterprise agents
The new open multimodal model targets continuous background tasks on Nvidia hardware, with smaller variants for local devices.
Key points
- Nemotron 3.5 Super VL has 120 billion parameters but activates only 12 billion per token.
- Two Bolt models target local machines with 24 GB and 4 GB of memory respectively.
- Nvidia claims the model matches its larger Ultra variant in accuracy on agent tasks.
Nvidia has released Nemotron 3.5 Super VL, an open multimodal model designed for enterprise agents that run continuously in the background. According to The Decoder, the model contains 120 billion parameters but uses a Mixture-of-Experts architecture where only 12 billion parameters are active per token. This design aims to reduce computational overhead while maintaining performance on document analysis and text processing tasks within organizational environments.
What is the performance profile?
Nvidia states that Super VL matches the accuracy of its larger Nemotron 3 Ultra model on specific agent tasks, despite having only one-fifth of the parameters. According to Nvidia, the model outperforms Ultra on Hermes and LangChain agent tests. It also surpasses its predecessor, Nemotron 3 Super, on coding benchmarks including SWE-bench and Terminal-Bench, as reported by The Decoder.
How is it optimized for hardware?
All Nemotron variants are optimized specifically for Nvidia hardware to drive chip and system sales, according to The Decoder. Nvidia claims that on a single GPU, Super VL delivers up to five times the throughput of other open models of similar size on agent tasks. This performance metric cites the Artificial Analysis leaderboard, suggesting significant efficiency gains for organizations deploying these models on Nvidia infrastructure.
What other models are available?
Two smaller Bolt models launch alongside Super VL to bring agents to local devices. Nemotron 3.5 Bolt 74B-A9B 24GB runs on systems with 24 GB of memory or more. A smaller Bolt E4B 4GB version targets fast voice assistants and will follow later this month. Both offer high intelligence per gigabyte of memory, according to Nvidia, enabling edge computing scenarios without cloud dependency.
Questions to ask your suppliers
Teams should verify whether suppliers are using open models with known hardware dependencies that could lock in infrastructure costs. Check if the supplier’s agent workflows require continuous background processing that might impact network security policies. Determine who within the organization monitors model behavior for unintended data extraction. Ensure that data retention policies align with the model’s document analysis capabilities. Identify which internal teams need visibility into third-party AI tool usage. Confirm that suppliers disclose all training data sources to prevent compliance risks.
What to do and how to stay safe: Nvidia
- Review infrastructure requirements to ensure hardware compatibility with new model deployments.
- Audit data handling procedures for any agent-based systems processing sensitive documents.
- Monitor vendor communications for updates on model patches or security advisories.
- Evaluate current security controls against the potential for automated tool use by AI agents.
Step-by-step guide: Browser Security Settings for Small Business Teams
General security guidance from the TechSec Feed newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
What is the size of the Nemotron 3.5 Super VL model?
It has 120 billion parameters, with 12 billion active per token.
Does it require Nvidia hardware?
Yes, all variants are optimized for Nvidia hardware to drive sales.
Are there versions for local devices?
Yes, two Bolt models target 24 GB and 4 GB memory systems.
