Nvidia wants AI agents to handle more complex work without requiring every task to run through a massive cloud-hosted model.
The company has introduced three Nemotron models, led by the 120-billion-parameter Nemotron 3.5 Super VL, which can analyze documents, interpret images, and use external tools to complete multistep tasks. Two smaller Bolt models are designed to bring similar agent capabilities to compatible PCs and edge devices.
Unlike models focused primarily on text generation, Super VL can extract information from PDFs, interpret charts and tables, and use tools to execute multistep workflows. That combination could help automate document-heavy work in finance, cybersecurity, engineering and research.
The model is available through Hugging Face, OpenRouter, ModelScope and Nvidia’s NIM deployment platform. Organizations can also fine-tune it using proprietary data, tools and policies.
Two smaller models bring agents closer to users
Nvidia also introduced two Nemotron 3.5 Bolt models designed to run AI agents locally. The Bolt 74B-A9B uses 2-bit quantization, a compression technique that reduces the memory needed to store model weights. Nvidia says this allows the 74-billion-parameter mixture-of-experts model, which activates about nine billion parameters at a time, to fit within a 24GB memory footprint.
The model targets local research, web browsing, coding, and debugging on compatible Nvidia RTX and DGX systems.
The model is available now through Hugging Face, with support for tools including Ollama, LM Studio, llama.cpp and vLLM.
The smaller Bolt E4B requires 4GB of memory and targets conversational agents, voice assistants, interactive characters and physical AI applications. It is designed for GeForce RTX, Jetson and RTX/DGX Spark hardware, with release scheduled for Oct. 23.
Together, the three models give developers options for deploying agents across different computing environments rather than relying exclusively on remote AI services.
Open models give enterprises more control
Nvidia trained Super VL on about 26 trillion tokens, then post-trained it on about 3.8 million samples, followed by reinforcement learning across dozens of environments. The company is releasing model weights, datasets and training recipes to support customization.
Early adopters include Scale AI, which post-trained the model for Skydance’s vendor credit finance agent and EY’s merger-and-acquisition due diligence agent. Other organizations are adapting it for cybersecurity, financial document analysis, coding and medical claims processing.
Nvidia also says its Dynamo serving platform can support up to eight times as many users while reducing latency by 85% under specified deployment conditions.
Local AI still depends on hardware
The Bolt models could make private, locally run agents more practical for developers and businesses handling sensitive information. Keeping workloads on compatible hardware can reduce dependence on cloud services, though privacy still depends on system configuration, application permissions and how connected tools handle data.
The 24GB model’s compressed format also comes with an important caveat: fitting a model into memory does not automatically guarantee the best accuracy or responsiveness for every task. Businesses will need to test reliability, speed and resource use against their own workloads before deploying agents at scale.
More must-read AI coverage
- SS&C Intralinks DealCentre AI vs. Datasite: Which platform is built for the future of dealmaking?
- SS&C Intralinks FundCentre AI vs. Juniper Square: Which platform better supports modern private markets fund managers?
- Why Data, Not Models, Determines AI Success
- The Rise of the AI-Native Factory: How Physical AI Is Transforming Manufacturing
What Nvidia’s new models mean for developers and businesses
The Bolt models could make locally run AI agents more practical for developers and businesses that want greater control over their workloads. Running models on compatible hardware can reduce dependence on cloud services, although privacy still depends on system configuration, application permissions, and how connected tools handle data.
For developers, Bolt could support coding assistants and research agents that operate on a compatible PC without sending every request to a cloud provider. Meanwhile, Super VL offers businesses another option for building specialized agents around internal documents and workflows.
However, smaller memory requirements do not automatically guarantee better performance. Organizations will still need to evaluate model accuracy, responsiveness, hardware compatibility, and resource use before deploying these systems at scale.
Consumers may eventually encounter smaller models in local voice assistants, interactive game characters, and device interfaces. But the immediate significance of Nvidia’s launch is the broader range of hardware and deployment options available to developers building AI agents.
Other news: Nvidia has made its entire Run:ai GPU orchestration platform open source, allowing organizations to manage GPU resources and AI workloads on Kubernetes without a software subscription.