Nvidia has introduced three new artificial intelligence models designed to power smarter digital assistants across business servers, personal computers and smaller devices, expanding its push to make advanced AI available beyond large cloud data centers.
They announced the Nemotron 3.5 models earlier this week, introducing a powerful system called Nemotron 3.5 Super VL alongside two smaller models in a new Bolt family. The first two models are available now, while the third is scheduled for release later this month.
The new lineup focuses on AI agents, software that can do more than answer questions. These systems are designed to perform tasks on a user’s behalf, such as analyzing documents, researching information, writing computer code and working with other applications.
Nvidia aims to give businesses and developers more flexibility in where they run AI, how much they spend and how they handle sensitive information.
Three Models Designed for Different AI Tasks
The most powerful model in the announcement, Nemotron 3.5 Super VL, is built for complex business activities that require understanding both written and visual information.
It can examine documents, read charts, interpret tables and analyze images while following instructions and using software tools. For example, an AI assistant built with the model could review a financial report, identify important figures and prepare an analysis using information from several pages.
Super VL contains 120 billion parameters, the internal settings that determine how an AI model processes information and generates answers. However, it uses only about 12 billion at a time, reducing the computing power needed to perform tasks.
According to Nvidia’s technical explanation, Super VL also supports extended conversations and large collections of information, making it suitable for tasks that involve multiple documents or lengthy research.
The second model, Nemotron 3.5 Bolt 74B-A9B 24GB, is designed to bring similar AI agent capabilities to compatible personal computers and workstations.
It can fit into 24GB of memory, allowing it to operate on supported Nvidia RTX and DGX Spark systems. Nvidia says the model can help users research topics, develop software and troubleshoot code without sending the AI processing to a remote server.
That could appeal to businesses handling confidential documents or developers who prefer running AI software locally. However, keeping a model on a computer does not automatically guarantee privacy, particularly if connected applications still exchange information with online services.
The third model, Nemotron 3.5 Bolt E4B 4GB, is designed for devices with more limited computing resources. Nvidia is targeting applications such as voice assistants, interactive digital characters and embedded AI interfaces. It is expected to run on supported GeForce RTX and Jetson hardware.
While Super VL emphasizes advanced reasoning and document analysis, the smaller Bolt models prioritize efficiency, allowing developers to choose a system suited to their available hardware and intended tasks.
Independent Tests Show Speed Gains, but Accuracy Remains a Challenge
Nvidia is promoting improved performance as a major advantage of the new models, and independent evaluations offer some support for those claims.
In an assessment, AI benchmarking company Artificial Analysis gave Super VL a score of 26.5 on its Intelligence Index, compared with approximately 13 for the earlier Nemotron 3 Super model.
Artificial Analysis also found that Super VL generated text roughly five times faster than Thinking Machines’ Inkling Small in a test involving lengthy prompts. The results were measured using a prerelease cloud service, so performance may differ when the model runs on other hardware.
The smaller Bolt 74B model also performed competitively in Artificial Analysis testing, scoring 21.4 on the same intelligence benchmark despite its lower memory requirements.
Faster responses don’t always translate into more reliable answers. In a separate evaluation by data visualization company Plotly, Super VL correctly answered 30 of 43 questions involving data analysis, achieving a 70% success rate.
Although it was the fastest of 28 models on Plotly’s leaderboard, answering a typical question in approximately 12 seconds, it failed to identify several deliberately introduced problems, including missing data and hidden patterns.
Those shortcomings matter for businesses considering AI assistants for financial reporting, operations or other activities where unnoticed errors could affect decisions. The results also show why organizations need to evaluate models against their own requirements rather than relying exclusively on advertised performance figures.
Nvidia Pushes Customizable AI for Businesses
Beyond speed, Nvidia is emphasizing the ability to customize its new models for specific business needs.
Unlike AI systems available only through commercial cloud services, Nemotron’s open-weight approach allows developers to download and modify the models’ trained parameters, subject to the applicable license. Organizations can also deploy the software on infrastructure they control.
That flexibility can be useful for companies that want AI assistants trained to understand their documents, internal procedures or specialized industry terminology.
Nvidia says several businesses are already adapting Super VL for professional applications. Scale AI has customized the model for financial workflows, while Ramp is exploring its use in processing financial documents. Other companies, including CrowdStrike, Atlassian and Salesforce, are working with the technology for cybersecurity, workplace software and government-related applications.
Early testing suggests that specialized training can improve performance on particular tasks.
Researchers at Thoughtworks AI Labs trained a prerelease version of Super VL to better interpret detailed charts. Accuracy on a specific chart-reading test increased from 17% to approximately 78% after several hours of additional training.
The specialized training also reduced accuracy on some other chart-related questions. The tests used a prerelease version of Super VL, and the results have not been confirmed on the final model.
Such trade-offs are particularly important when companies adapt general-purpose AI systems for specialized work.
For Nvidia, the expanded Nemotron family also supports its broader strategy of providing both the hardware and software used to run AI applications. More efficient models could allow customers to run advanced AI applications on existing hardware, reducing their dependence on expensive cloud infrastructure.
Super VL is available through platforms including Hugging Face, OpenRouter and Nvidia’s own AI deployment services. The 24GB Bolt model is also available on Hugging Face, with support for popular local AI tools such as Ollama and LM Studio.
Leave a Reply