NVIDIA Expands Open AI Model Lineup With Nemotron 3.5 Super VL and Bolt
The new models target enterprise AI agents and local computing, extending NVIDIA’s push to make advanced AI systems more customisable and deployable across different hardware environments.
NVIDIA has expanded its Nemotron family of artificial intelligence models with the release of Nemotron 3.5 Super VL, a 120-billion-parameter multimodal model, alongside two smaller Bolt variants designed for local computing. The launch, announced on October 15, 2026, aims to give developers more options for building AI agents that can interpret visual information, use software tools and execute multistep tasks across enterprise and edge environments.
The announcement reflects a growing focus within the AI industry on systems that do more than generate text. Such agents are designed to interpret instructions, interact with digital tools and complete sequences of tasks with limited human intervention. NVIDIA’s latest models are intended to support these capabilities while giving organisations greater control over deployment and customisation.
A multimodal model for complex workflows
Nemotron 3.5 Super VL combines language reasoning, instruction following, tool use and visual understanding in a single model. According to NVIDIA, its capabilities extend to images, documents, charts and video, allowing developers to build applications that process different forms of information without necessarily relying on separate vision-language models.
The model uses a mixture-of-experts architecture with 120 billion parameters, of which 12 billion are activated for each token processed. This design allows the system to retain a large pool of specialised model capacity while activating only a portion of it for individual computations.
NVIDIA says Super VL delivers up to five times the throughput of competing open models in its size category on agent-oriented tasks. That figure is a company-reported performance claim rather than a universal measure of superiority; actual results depend on the workload, hardware, evaluation method and competing models selected.
The intended applications include document analysis, software development, visual data extraction and business processes involving multiple steps. For enterprises, combining these functions within one model could simplify system architecture and reduce the need to transfer information between separate AI components.
NVIDIA has also positioned the model as a foundation for specialised agents rather than a fixed, general-purpose assistant. Developers can further train it using their own data, tools and operational requirements, potentially adapting its behaviour to particular industries or internal workflows.
Bolt brings AI agents closer to local hardware
Alongside Super VL, NVIDIA introduced two Bolt models intended to extend agent capabilities beyond large data-centre deployments.
Nemotron 3.5 Bolt 74B-A9B is designed for systems with 24GB of memory, using low-bit quantisation to reduce its memory footprint. NVIDIA identifies local RTX and DGX Spark systems, among other supported configurations, as potential deployment environments. The model is intended to balance performance and resource consumption for tasks that can run on local hardware.
The second variant, Nemotron 3.5 Bolt E4B, is designed around a 4GB footprint. It targets lightweight conversational and edge applications, where available memory and computing resources are more constrained.
Super VL and Bolt 74B-A9B were available at announcement, while NVIDIA scheduled Bolt E4B for October 23. The company said the models would be distributed through relevant model platforms and its own AI software ecosystem, depending on the variant.
The distinction between the models is practical. Large models can support more demanding reasoning and multimodal tasks, but they require substantial computing resources. Smaller, quantised models can bring selected capabilities to local devices, potentially reducing dependence on remote inference services.
Local execution can also give organisations more control over where data is processed. However, it does not automatically guarantee privacy or security: those outcomes depend on the software configuration, access controls, data handling and deployment environment.
Open models and the competition for developers
The release also advances NVIDIA’s broader effort to strengthen its position in the AI software ecosystem. Its business extends beyond supplying processors to include inference software, development frameworks and tools for deploying AI applications.
Open-weight models offer developers access to model parameters and greater flexibility than many closed commercial systems. Depending on the licence and accompanying resources, organisations can adapt models, host them on their own infrastructure and integrate them into existing applications.
NVIDIA’s Nemotron strategy therefore addresses more than model performance. It seeks to make the company’s hardware and software platforms attractive throughout the development process, from training and customisation to inference and production deployment. The availability of models suited to different computing environments could help developers build applications without having to adopt a single deployment model for every task.
Nevertheless, open availability does not eliminate the costs of running AI systems. Enterprises must still account for computing infrastructure, electricity, engineering work, model evaluation and ongoing maintenance. Performance claims also need to be tested against the workloads an organisation actually intends to run.
Performance claims will need independent scrutiny
NVIDIA’s emphasis on throughput and efficiency addresses a central challenge in agentic AI: an agent may make repeated model calls while gathering information, using tools and checking intermediate results. Improvements in inference speed can therefore have a meaningful effect on the cost and responsiveness of complex workflows.
But throughput alone does not establish whether a model is the best choice for a particular task. Accuracy, reliability, latency, memory requirements and the ability to recover from errors can be equally important, especially when agents operate with limited supervision.
Independent testing across comparable hardware and representative workloads will be necessary to establish how Super VL and Bolt perform against alternative models. Organisations will also need to determine whether the models’ customisation options translate into measurable improvements in their own operations.
NVIDIA’s latest release broadens the range of AI agents developers can build across data centres, workstations and edge systems. Its longer-term impact will depend less on the number of models it introduces than on whether those models can deliver dependable results at a competitive operating cost.
