NVIDIA Corporation (Nasdaq: NVDA) has expanded its Nemotron model family with an open 120-billion-parameter multimodal model for large enterprise agents and two smaller models designed for local and edge deployments. The additions give organizations another option when they want to adapt and run agent models on their own infrastructure instead of relying only on a hosted frontier-model service.
The largest release, Nemotron 3.5 Super VL, is intended for agents that must handle text, images, web pages, PDFs, tables and charts as part of a business workflow. NVIDIA said the 120-billion-parameter model can perform instruction following, tool use, visual understanding and multistep work, while operating at roughly one-fifth the parameter count of its largest reasoning model. Those are vendor performance claims, but the design focus is relevant to enterprise workloads built around documents and systems of record rather than simple chat prompts.
For a procurement, finance, engineering or security workflow, visual and structured-document understanding can reduce the need to shuttle a task among separate optical-character-recognition, extraction and language-model services. NVIDIA said Super VL carries context from a document or web page through to data and a decision. An enterprise still needs to validate the accuracy of that chain for its own documents, policies and exceptions before using it in a consequential process.
NVIDIA also released Nemotron 3.5 Bolt 74B-A9B 24GB, which it said uses native 2-bit quantization to fit in 24GB of memory and run alongside other applications on RTX and DGX Spark hardware. The company positions it for local research, coding and other agent work that an organization may prefer to keep outside a public cloud. A smaller Bolt E4B 4GB model targets fast conversational and embedded agents on GeForce RTX, Jetson and related systems.
The deployment range is the material distinction. Many enterprises are deciding workload by workload whether data, latency, cost or residency requirements justify a centralized cloud service, a private data center, an endpoint or an edge device. NVIDIA is offering one model family across those environments, from data-center and cloud systems to workstations and embedded hardware. That can simplify model evaluation, but it does not remove the operational differences among those locations.
The company said the models ship with open weights, datasets and training recipes, and can be customized through NVIDIA NeMo. It cited use cases including agentic search, engineering design, financial-document analysis, cybersecurity research and coding agents. Open assets can give an organization more control over adaptation and deployment; they also leave the organization responsible for evaluating data provenance, model behavior, access controls and change management.
NVIDIA NIM packages Super VL as a production microservice with validated configurations and enterprise support, according to the company. It also said the model can run with NVIDIA Dynamo, which the company claims can increase served users and lower latency. Organizations should benchmark those claims against their selected hardware, model context lengths and concurrency requirements rather than treating them as general outcomes.
Compared with the legacy choice between a closed hosted model and a single specialized local model, the Nemotron releases move more capability into a configurable deployment spectrum. The strategic value is not a universal model winner; it is the ability to align an agent’s model, location and control boundary with the workload it is meant to operate.