Nvidia's 120B MoE Nemotron 3.5 runs local agents via compact 4GB to 24GB Bolt variants
The 120-billion parameter open model targets background agent workflows and local hardware deployment via compact Bolt variants.
Key points
- Nvidia released Nemotron 3.5 Super VL, an open multimodal model with 120 billion parameters for enterprise agents.
- Two Bolt variants target local hardware, including a 24GB version for agents and a 4GB version for voice assistants.
- The model uses a Mixture-of-Experts architecture, activating only 12 billion parameters per token.
Nvidia has launched Nemotron 3.5 Super VL, an open multimodal model designed for continuous enterprise agent operations. According to The Decoder, the 120-billion parameter model handles text, images, documents and videos. It targets background processing tasks within organisational workflows. The release includes training recipes and datasets alongside the model weights.
Model Architecture
Nvidia developed the model using a Mixture-of-Experts architecture. As detailed in a technical blog post, only 12 billion of the total parameters activate per token. The company distilled Super VL from Nemotron 3 Ultra, its largest reasoning model. This approach aims to match Ultra's accuracy while using one-fifth the parameters.
Performance Claims
The model is built for specialised agents that run continuously. It processes multi-step business processes, tool calls and document analysis. Nvidia claims it scores higher than Ultra on agent tests from Hermes and LangChain. It also beats its predecessor on coding benchmarks SWE-bench and Terminal-Bench.
Available Variants
- Nemotron 3.5 Super VL (120B parameters)
- Nemotron 3.5 Bolt 74B-A9B 24GB (local agents)
- Nemotron 3.5 Bolt E4B 4GB (voice assistants)
Deployment Considerations
- Verify hardware compatibility with Nvidia systems for optimal performance.
- Review training recipes provided alongside model weights.
- Assess local memory requirements for Bolt variants.
- Monitor benchmark results against existing internal models.
What it means for organisations
Teams should evaluate whether continuous agent workflows align with current operational needs. Infrastructure managers must confirm hardware compatibility, as the models are optimised for specific vendor systems. Data governance teams should review the provided training recipes for compliance. Security leads need to assess how background agent processing interacts with existing access controls. Procurement should consider the cost implications of dedicated hardware requirements versus cloud alternatives.
What to do and how to stay safe: Nvidia
- Audit current agent workflows for compatibility with multimodal inputs.
- Confirm local hardware meets memory requirements for Bolt variants.
- Review dataset sources included in the release for compliance.
- Test throughput on existing infrastructure before full deployment.
General security guidance from the TechSec Observer newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
What is the parameter count of Nemotron 3.5 Super VL?
The model contains 120 billion parameters, with 12 billion active per token.
Are there smaller versions for local devices?
Yes, Bolt variants include a 24GB model for agents and a 4GB version for voice assistants.
Who helped develop the Nemotron 3.5 Super VL model?
Nvidia developed it with the Nemotron Coalition, including Mistral AI, Perplexity and Black Forest Labs.
