Nvidia Wants to Put AI Agents Everywhere, From Data Centres to Your PC

Nvidia has expanded its artificial intelligence model portfolio with three systems designed to run complex enterprise workflows in data centres and power AI agents directly on personal computers, intensifying its push to extend its influence beyond chips into the software that businesses use to deploy AI.

The company today introduced Nemotron 3.5 Super VL, a 120-billion-parameter model for visual understanding and automated business tasks, alongside two smaller models designed to operate within the memory constraints of local computers and edge devices.

The announcement, issued by Kari Briski, Nvidia’s vice president of generative AI software for enterprise, highlights the company’s strategy of offering businesses a range of models that can be customised and deployed across its hardware ecosystem.

Nvidia said Super VL could deliver accuracy comparable with its largest reasoning model while using roughly one-fifth the number of parameters. It also claimed the model completed tasks up to five times faster than the next-best model in its class, although that figure depends on the benchmarks and testing conditions used.

“Any organization that wants to run agents on NVIDIA infrastructure now has an open, optimized model for their tier, without having to choose between capability and deployability,” Briski said.

The launch reflects a broader shift in the AI industry towards agentic systems, software that can interpret information, use digital tools and execute multistep tasks rather than simply respond to questions.

For Nvidia, the opportunity is to make its computing platforms more attractive by supplying not only the processors that run AI workloads but also the models, development tools and inference software that businesses need to put those systems into production.

Open-weight AI gains ground

Nvidia’s expansion comes as open-weight models gain traction among businesses seeking greater control over their AI deployments, including the ability to customise systems, operate them on their own infrastructure and reduce dependence on proprietary services.

Vercel’s September 2026 AI Gateway Production Index, which analysed traffic through August, found that open-weight models accounted for 56% of tokens processed through its gateway, up from 7% in December 2025. The average token cost had also fallen by more than half over five months, according to the report.

The figures reflect Vercel’s platform rather than the entire AI industry, but they illustrate the growing role of models that can be downloaded and deployed outside a single provider’s hosted service. 

The commercial attraction is particularly strong for organisations running agents that repeatedly examine documents, search databases, call software tools and process information. Each additional operation consumes computing resources, making inference efficiency an important consideration as businesses scale their AI deployments.

Nvidia’s Nemotron family is designed to address different parts of that market. Lightning targets high-throughput workloads, Ultra is intended for demanding reasoning tasks, Super VL focuses on multimodal and agentic work, while Bolt is aimed at local and resource-constrained environments.

The strategy could help Nvidia capture more of the software spending associated with AI adoption. However, businesses will still need to compare model performance, operating costs, reliability and compatibility with their existing systems before committing to a particular platform.

Super VL targets complex business workflows

Nemotron 3.5 Super VL is designed to interpret text and visual information, including webpages, PDFs, charts, tables and other structured documents, while supporting agents that use tools and perform multistep tasks.

Nvidia said the model was trained on approximately 26 trillion tokens across pretraining and continued pretraining, followed by post-training using roughly 3.8 million samples and reinforcement learning across dozens of environments.

The company is releasing model weights, datasets and training recipes, allowing developers to adapt the model to their own data, business rules and workflows. Its NeMo development tools support customisation, while Nvidia NIM packages the model for production deployment.

The approach could be useful in financial services, where agents may examine credit agreements or identify suspicious claims; in cybersecurity, where systems need to analyse threats; and in engineering, where software must interpret technical documents and designs.

Nvidia named several organisations working with Super VL, including Scale AI, which post-trained the model for a vendor credit finance agent for Skydance and an mergers-and-acquisitions due-diligence agent for EY. Hebbia is using a customised version for credit-agreement analysis, while Mercor has worked on coding-agent capabilities.

Atlassian, CrowdStrike, Dassault Systèmes, Perplexity, Salesforce and Synopsys were also identified as software companies integrating or customising the model for their respective applications.

These deployments demonstrate the range of potential use cases, but the announcements do not independently establish the models’ commercial performance or prove that they outperform competing systems in production.

Smaller models bring AI closer to users

The Bolt models extend Nvidia’s strategy towards local AI, where inference takes place on a user’s computer or another nearby device instead of relying entirely on a remote data centre.

Nemotron 3.5 Bolt 74B-A9B has 74 billion total parameters, with approximately nine billion activated for each token under its mixture-of-experts design. Nvidia says native two-bit quantisation allows the model to fit within 24GB of memory.

The model is designed to run on compatible Nvidia RTX systems and DGX Spark hardware, supporting tasks such as web research, writing and code debugging through local agent frameworks.

Nvidia said the model ranked first among open models in its class on PinchBench and second among open models on BrowseComp. Those rankings are company-reported claims and should be assessed against the benchmarks’ evaluation methods and competing models’ results.

The smaller Bolt E4B model is designed to fit within four gigabytes of memory and targets conversational agents, voice assistants and embedded applications. Nvidia said it would be available on October 23.

Local deployment offers potential advantages for businesses that want to keep sensitive documents within their own infrastructure, reduce dependence on internet connectivity or avoid recurring charges for every request to a hosted model.

It does not, however, automatically guarantee privacy or security. Applications still need appropriate access controls, secure storage, software updates and safeguards against unauthorised data access.

Memory requirements also represent only part of the cost of operating AI. Processing speed, electricity consumption, hardware availability and the quality of the model’s responses remain important factors in determining whether local deployment is economical.

DeepSeek and other rivals raise the pressure

Nvidia faces competition from companies pursuing different approaches to making AI models more capable and economical.

DeepSeek has attracted attention for its emphasis on efficient model development and relatively low-cost inference. Alibaba’s Qwen family, Meta’s Llama models and Google’s Gemma portfolio provide additional options for developers seeking models they can adapt and deploy.

Vercel’s September report found that open-weight models had captured a majority of token volume on its gateway, underscoring the growing competitive pressure on providers to deliver capable models at lower costs. The same report showed that pricing and usage patterns can change rapidly as new models enter production. 

Meanwhile, proprietary providers such as OpenAI, Anthropic and Google continue to compete through hosted models, developer platforms and managed services. These may appeal to organisations that prioritise operational simplicity, integrated support or particular capabilities over direct control of model weights.

Nvidia’s distinction is its integration of models with its own hardware, inference software and development tools. That combination could simplify optimisation for customers already invested in Nvidia systems, but it also raises questions about how easily businesses can move workloads to other chips or cloud providers.

Open weights offer a degree of flexibility, but they do not by themselves ensure hardware portability. Performance may still depend on specialised software, supported operators, quantisation methods and the underlying processor architecture.

For enterprises, the key question is therefore not simply which model tops a benchmark. It is which combination of model, hardware and software delivers the required accuracy and reliability at an acceptable total cost.

The test is whether enterprises can scale adoption

Nvidia said Super VL was being customised for applications ranging from financial analysis and cybersecurity to scientific discovery and medical claims processing.

The company is making the model available through Hugging Face, OpenRouter, ModelScope and its NIM deployment service. Bolt 74B-A9B is also available through Hugging Face, with support for local inference frameworks including llama.cpp, Ollama, LM Studio and vLLM.

Nvidia further claimed that running Super VL through its Dynamo inference software could allow organisations to serve up to eight times more users with 85% lower latency. Those figures are vendor claims; their practical significance depends on the workloads, hardware configurations and baseline against which they were measured.

The release highlights a central challenge for the AI industry: moving from impressive demonstrations to systems that can operate reliably and economically at scale.

For Nvidia, expanding Nemotron gives developers more options for deploying AI across data centres, workstations and edge devices while reinforcing the value of its wider computing platform.

But competition is increasingly centred on more than raw model size or speed. Companies must demonstrate that their systems can complete useful tasks accurately, integrate into existing workflows and generate measurable returns on the cost of computing.

Nvidia’s latest models expand the range of tools available to developers. Their longer-term significance will depend on whether independent evaluations and real-world deployments confirm the company’s claims — and whether businesses choose Nemotron over the growing number of alternatives.

Stay ahead of the Stories shaping our world. Subscribe to Impact Newswire and join our 
WhatsApp Channel for updates on global tech, business, and innovation—all in one place.

Dive deeper into the future with the Cause Effect 4.0 Podcast, where we explore the ideas, trends, and technologies driving the global AI conversation.

Got a story to share? Contact Us to reach a global audience with Impact Newswire.