Nvidia launches Nemotron 3.5 Super VL and Bolt models for agents
New open-weight AI models offer high-speed agentic capabilities for enterprise servers and local edge hardware deployments.
Key points
- Nemotron 3.5 Super VL is a 120B parameter multimodal model optimized for speed and accuracy.
- Two Bolt variants enable local execution on workstations and edge devices with reduced memory needs.
- Open weights allow organizations to post-train and deploy models behind their own firewalls.
Nvidia announced the release of customizable Nemotron 3.5 Super LV large language models and edge-focused 3.5 Bolt variants for local systems, according to SiliconANGLE. The company stated these three new models address the full spectrum of agent deployments, ranging from large enterprise servers down to local hardware.
The Nemotron 3.5 Super VL is an open 120-billion-parameter model developed with the Nemotron Coalition, designed for high-accuracy multimodal agentic tasks. Nvidia claims this model delivers power comparable to its largest reasoning models while using only one-fifth of the size. This efficiency allows Super VL to generate outputs nearly five times faster than other models in the same class.
Model Architecture and Training Data
Nvidia trained Super VL using open data and methods, processing approximately 26 trillion tokens during pretraining and continued pretraining phases. The process included 3.8 million post-training samples and reinforcement learning across 100 open environments to fine-tune the model’s capabilities. The company also released open datasets on Hugging Face covering agentic tool use, finance, safety, fairness, and vision-language understanding.
Because the weights are open, organizations can post-train Super VL on proprietary data and deploy it on private infrastructure. This approach allows enterprises to keep the model behind their firewall, maintaining data sovereignty, privacy, and safety without relying on closed frontier models.
Local and Edge Deployment Options
The Bolt 74B-A9B 24GB model is a smaller variant based on Super VL, designed to run agentic tasks locally. It employs a hybrid mixture-of-experts architecture that activates 9 billion of its 74 billion parameters at a time. Using two-bit quantization, the model fits within 24 gigabytes of memory, enabling it to run on enterprise workstations or PC graphics processing units.
Users can delegate everyday projects, such as writing and debugging code, to this model without sending workloads to the cloud. Nvidia also announced an even smaller model, Bolt E4B 4GB, launching later this month. This tiny model is built for edge deployment on GeForce RTX, Jetson, and RTX/GTX Spark hardware.
What should you do?
Security and technology professionals should evaluate whether local model deployment aligns with their data governance policies. Review memory and hardware requirements for the Bolt variants to ensure compatibility with existing workstation or edge infrastructure. Monitor the open datasets and model weights on Hugging Face for updates or documentation changes. That suggests organizations can begin planning post-training workflows using their own data sets.
Background: Large language models
Large language models are a specific type of neural network trained on text. Generative AI is the broader umbrella term for any system that creates new content, including images and code. Not all generative AI uses large language models, but most modern text generators do.
What to do and how to stay safe: Nvidia
- Assess current hardware capabilities to determine if local GPU resources can support 24GB or smaller memory footprints.
- Review data privacy policies to understand the implications of running inference locally versus cloud-based services.
- Monitor the Nemotron Coalition and Nvidia for updates on model performance benchmarks and security patches.
- Evaluate the open datasets for relevance to specific industry use cases like finance or code debugging.
Step-by-step guide: Prompt Injection Warning Signs: How to Spot AI Manipulation
General security guidance from the Threat Times 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?
The Nemotron 3.5 Super VL model has 120 billion parameters.
How much memory does the Bolt 74B-A9B model require?
The Bolt 74B-A9B model is quantized to fit within 24 gigabytes of memory.
When will the smallest Bolt model be available?
The Bolt E4B 4GB model is scheduled to launch later this month.
