A large model for enterprises, lighter versions for local use
Nvidia has announced new open-weight Nemotron models for enterprise agents and personal computers. TechCrunch reports that Nemotron 3.5 Super VL has 120 billion parameters and is designed for tasks requiring tool use and multimodal capabilities. The company also introduced two smaller Bolt models, intended for AI workstations and PCs: one with 74 billion parameters and an E4B version for conversational agents.
Availability of the weights allows organizations to run and adapt a model on their own infrastructure, in compliance with the relevant licenses. Nvidia claims that Super VL can approach the accuracy of larger models on instruction and tool-use tasks, and that it outperforms some competitors after task-specific training. These are company-reported results and should be read in light of the chosen benchmarks, test data, and configuration used.
The advantage of local execution and its costs
Running a model on the device can reduce the need to send documents and requests to a cloud service and can also work over limited connections. But a local model is not automatically private or secure: device protection, updates, data control, and management of the agent’s permissions are required. Hardware requirements also matter; the ability to run a model depends on memory, accelerators, optimization, and the size of the tasks.
Nvidia says the Bolt models can support agentic tasks such as multi-step reasoning, tool calls, and web navigation. However, each added capability expands the risk surface: an assistant that can use browsers or files must be constrained and supervised, especially when it operates on corporate or personal data.
A strategy that combines models and hardware
The presentation reinforces Nvidia’s strategy beyond selling GPUs. Offering open models optimized for its own systems can make it easier to build applications within the company’s ecosystem. The company lists enterprises including Salesforce, EY, Ramp, CrowdStrike, and Perplexity among users of the larger model; according to Nvidia, the list indicates adoption or experimentation, but it does not amount to publicly comparable results for every case.
For enterprises, the decision is not simply “open versus closed.” License, adaptation costs, performance on their own task, availability of support, and hardware dependence all matter. Evaluating the models requires reproducible tests, including difficult cases and error measures, not just benchmarks selected by the vendor.
Super VL and Bolt 74B-A9B are listed as available, while Bolt E4B is expected later in the month according to the report. The move makes models oriented toward agents more accessible, but their real value will depend on how well they perform in everyday applications and what controls are put in place around them.
