Nvidia's 120B Nemotron Super VL powers enterprise agents with 4GB Bolt variants
Nvidia launches an open multimodal model with 120 billion parameters designed for continuous background agent tasks and document analysis in corporate environments.
Key points
- Nemotron 3.5 Super VL uses a Mixture-of-Experts architecture with 120 billion parameters.
- Two compact Bolt variants target local hardware with 24 GB and 4 GB memory requirements.
- The model is optimized for Nvidia hardware and developed with the Nemotron Coalition.
Nvidia has released Nemotron 3.5 Super VL, an open multimodal model containing 120 billion parameters that is engineered specifically for enterprise agent workflows and detailed document analysis.
The background
Nvidia designed this system to operate continuously in the background, processing complex inputs including text, images, documents, and video streams. The company states that the model matches the accuracy of its larger Nemotron 3 Ultra reasoning engine while using only one-fifth of the parameters. According to Nvidia, the new model scores higher than Ultra on agent tests from Hermes and LangChain.
Target audience
The release targets organizations deploying specialized agents that must follow instructions, call external tools, or execute multi-step business processes. Two smaller Bolt variants are also launching to bring these capabilities to local machines and edge devices. The Nemotron Coalition, which includes Mistral AI, Thinking Machines Lab, Perplexity, and Black Forest Labs, helped develop the model.
Cost and compliance
Nvidia claims its largest model delivers up to five times the throughput of other open models of similar size on a single GPU, according to the Artificial Analysis leaderboard. The company publishes datasets and training recipes alongside the weights but optimizes everything for its own hardware to drive chip and system sales.
What should you do?
Teams should evaluate whether their current infrastructure can support the memory requirements of the new Bolt variants. Managers need to verify if the model's multimodal capabilities align with specific enterprise agent tasks such as table extraction or chart interpretation.
The background in plain language
Enterprise agents are software programs that perform complex, multi-step tasks without constant human supervision. Unlike simple chatbots, these systems can interpret documents, call external tools, and maintain context over long interactions. Organizations must ensure these agents operate within strict security boundaries, as they often access sensitive data and execute actions on behalf of users.
What to do and how to stay safe: Nvidia
- Verify that your hardware meets the memory requirements for the specific Bolt variant you intend to deploy locally.
- Review the model's multimodal capabilities to ensure they align with your specific document analysis and agent task needs.
- Monitor official channels for the release timeline of the 4 GB Bolt variant targeting voice assistants.
- Assess whether your existing Nvidia hardware optimizations will provide the claimed throughput benefits for your workload.
Step-by-step guide: Mobile Security Architecture: 8 Practices for Device Integrity
General security guidance from the Vector Tribune 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, though only 12 billion are active per token due to its Mixture-of-Experts architecture.
How does Super VL compare to Nemotron 3 Ultra?
Nvidia states Super VL matches Ultra's accuracy with one-fifth the parameters and scores higher on specific agent tests from Hermes and LangChain.
What are the memory requirements for the Bolt variants?
The Bolt 74B-A9B requires 24 GB of memory, while the smaller Bolt E4B targets systems with 4 GB of memory.
