NVIDIA released new open-weight models on Thursday, designed to support always-on agents and edge computing.
Nemotron 3.5 Super VL is a large model with 120 billion parameters designed for enterprise-grade agents. NVIDIA states that, despite having fewer parameters than the Nemotron 3 Ultra model, it achieves comparable accuracy in instruction following and tool use tasks.
On the other end, NVIDIA also released two lightweight Bolt models designed to run locally on new AI PCs equipped with NVIDIA RTX Spark chips and on workstations running DGX Spark. These consumer-, developer-, and enterprise knowledge worker-oriented PCs are entering the market through partners such as Microsoft and Dell.
Kari Briski, Vice President of Open Models and Software at NVIDIA, wrote in the blog post introducing these models: “Any organization looking to run agents on NVIDIA infrastructure can now find an open, optimized model suited to its needs without having to choose between capability and deployability.”
NVIDIA stated that after fine-tuning for specific agent tasks, Super VL outperforms competing models from Anthropic and OpenAI (Opus 5, GPT 6.1 Sol, Luna, Astra) in task completion speed. These tasks include document intelligence, financial accounting, scientific experiments, health insurance claims processing, and security attack resilience.
Although such claims are common whenever major tech companies release new models, NVIDIA also lists a number of large companies allegedly using Super VL, including Skydance, EY, Ramp, CrowdStrike, Perplexity, and Salesforce.
Of course, this is not the first time NVIDIA has released open-weight models that can compete with proprietary cutting-edge labs—companies that are both its customers and its investors. NVIDIA’s Nemotron series of models encompasses functionalities such as multimodal AI, reasoning, programming, mathematics, and enterprise automation.
However, these rapidly growing proprietary AI labs are now turning to other chip and AI system providers beyond NVIDIA to meet demand, such as AWS’s Trainium, Cerebras, and in-house developed chips.
Therefore, NVIDIA is heavily betting on the trend of enterprises wanting to run open-source models in their own data centers. Briski wrote: “Unlike closed frontier models, organizations can fine-tune Super VL with their own data, deploy it on their own infrastructure, and own the results they build.”
But with the release of these smaller, locally runnable PC models, NVIDIA is also trying to advance the emerging trend of consumer-grade AI assistants. Although it seems that new AI assistants are being launched daily by startups and major tech companies, consumers concerned about privacy may not be willing to allow such agents deep access to their lives, including sensitive information about their homes, health, and finances.
Therefore, NVIDIA aims to create a new market by building machines capable of running models and agents locally, selling PCs powered by its chips. Briski writes that PC users “can delegate everyday tasks... without sending these workloads to the cloud.”
Among the two Bolt models, the Nemotron 3.5 Bolt 74B-A9B 24GB is a 74-billion-parameter model designed for local agent tasks, balancing speed and reasoning capability to handle multi-step reasoning, tool invocation, and web browsing. The smaller Bolt E4B is tailored for conversational agents.
NVIDIA has announced that Super VL and Bolt 74B-A9B are now available, with Bolt E4B set to launch later this month.