Nvidia introduced its new vulnerability-heavy models developed for corporate AI agents and AI applications that can run locally on personal computers. Nemotron 3.5 Super VL, announced by the company, aims to meet the needs of businesses for constantly working artificial intelligence agents, with its 120 billion parameter structure. In addition, two different Nemotron Bolt models that will work with lower hardware requirements are also offered to users. The new models target institutions that want to reduce dependence on cloud-based systems, as well as users who prefer to perform artificial intelligence operations directly on their own computers. Nvidia thus expands the Nemotron model family for both large-scale corporate infrastructures and local computer systems.
The Nemotron 3.5 Super VL can achieve similar levels of accuracy on certain tasks despite using fewer parameters than the company’s larger Nemotron 3 Ultra model. According to the information provided by Nvidia, the model offers results comparable to larger models, especially in executing commands correctly and using external tools. These features stand out for artificial intelligence agents that communicate with different software services and perform tasks consisting of multiple stages. In addition, the model’s adaptability to specific corporate workflows allows companies to develop customized systems with their own data. Thanks to its open-based structure, organizations can run the model on their own infrastructure and undergo additional training processes according to their needs.
Nvidia’s Vice President of Open Models and Software, Kari Briski, stated in his statement about the new models that organizations that want to run artificial intelligence agents on Nvidia infrastructure will have access to options suitable for different hardware levels. According to Briski, companies will not have to make a clear choice between the capabilities of the model and the hardware conditions of the environment in which it will be used. Nvidia draws attention to the flexibility offered by models that institutions can run in their own data centers. Unlike closed systems, it is possible for Nemotron 3.5 Super VL to be retrained with company data and used in infrastructures controlled by institutions. This approach provides an alternative for businesses that do not want to send sensitive data to third-party cloud services.
What do Nvidia Nemotron 3.5 Super VL and Bolt models offer?
Nvidia claims that the Nemotron 3.5 Super VL can produce results faster than some competing models from Anthropic and OpenAI when put through additional training on specific AI agent tasks. The company’s comparisons include different usage areas such as document analysis, financial accounting, scientific experiments, evaluation of health insurance claims and resistance to security attacks. However, it should be noted that these performance claims are based on test results shared by Nvidia. The success of models in real use conditions may vary depending on factors such as the hardware used, the complexity of the tasks and training data. Therefore, higher processing speed does not mean better results in every scenario.
The company also announced that there are names such as Skydance, EY, Ramp, CrowdStrike, Perplexity and Salesforce among the organizations using Nemotron 3.5 Super VL. The fact that companies from different sectors are included in the list shows that the model does not focus only on a single corporate usage area. Nvidia’s Nemotron family previously included models designed for tasks such as multi-modal artificial intelligence, reasoning, software development, mathematics and corporate automation. The new Super VL model also expands these options. On the other hand, adapting the model to institutions is not limited to the accessibility of weight files; It brings with it additional requirements such as data security, infrastructure management and regular evaluation of the model.
Other models in Nvidia’s announcement were the Nemotron Bolt options, designed to be run on personal computers and smaller-scale systems. Nemotron 3.5 Bolt 74B-A9B 24GB stands out as a model with a total of 74 billion parameters and designed for local artificial intelligence agent tasks. Nvidia aims for the model to strike a balance between processing speed and reasoning abilities. Supported usage areas include completing multi-stage tasks, calling a ride, and surfing the internet. The smaller Bolt E4B model was developed specifically for chat-based artificial intelligence agents that communicate with users via natural language.
Bolt models can be run natively on hardware such as artificial intelligence computers with Nvidia RTX Spark processors and DGX Spark-based workstations. It aims to make such applications reach a wider user base with computer systems offered by business partners such as Nvidia, Microsoft and Dell. Thanks to its local operation feature, artificial intelligence agents may not have to constantly send data to remote servers to perform certain tasks. This can be especially advantageous for users working with personal documents, financial information and other sensitive content. However, since the performance of local models depends on the memory capacity, processing power and software support of the computer, the same experience should not be expected on all systems.
The fact that artificial intelligence agents work on personal computers also offers a different option in terms of data privacy discussions. Today, many AI assistants can automate tasks by accessing users’ calendars, files, emails, and various online accounts. Although such access is useful, it can raise questions about sharing personal information with cloud services. Nvidia’s focus on native models allows certain operations to be completed on the device and data to be kept in systems under the user’s control. However, just because a model runs locally does not mean that all the applications it uses are offline or that no information is transferred to third-party services.
Nvidia’s emphasis on open-heavy models is also linked to competition in the artificial intelligence infrastructure market. The company’s graphics processors are widely used in training and running large artificial intelligence models. Despite this, major artificial intelligence developers are also evaluating alternative hardware such as AWS Trainium and Cerebras to reduce infrastructure costs and supply dependencies. Some technology companies are developing their own artificial intelligence processors. Expanding Nvidia’s Nemotron models allows it to adapt software and model options, as well as hardware products, to a wider variety of usage scenarios.
Nemotron 3.5 Super VL and Nemotron 3.5 Bolt 74B-A9B 24GB models are now available. Nvidia states that the smaller Bolt E4B model will be available later in the month when the announcement is made. The new options address different needs between high-capacity-demanding enterprise applications and lighter-weight AI agents running on personal computers. It is particularly noteworthy that models can be customized with institutions’ own data and certain operations can be performed on local hardware. On the other hand, the performance to be achieved in real use experience, hardware costs and maintenance requirements of the systems will be the main factors that determine in which areas these models will be preferred.