Nvidia launches new open models to run agents in the enterprise and on PCs

Nvidia announced on Thursday a set of new open-weight models aimed at always-on agents and edge computing, split between a large enterprise model and two small ones meant to run locally on AI PCs and workstations.
The headline release is Nemotron 3.5 Super VL, a 120-billion-parameter model built for enterprise agents. Nvidia says it uses fewer parameters than its Nemotron 3 Ultra model while delivering similar accuracy on instruction-following and tool-use tasks. Parameter count is a rough proxy for model size and serving cost, so a smaller model that holds accuracy on agentic work matters to teams paying for GPU capacity.
At the other end, Nvidia launched two lightweight Bolt models. Nemotron 3.5 Bolt 74B-A9B 24GB is a 74-billion-parameter model for local agentic tasks, positioned to balance speed with reasoning for multistep reasoning, tool calling and web browsing. The smallest, Bolt E4B, targets conversational agents. These are intended to run on the new AI PCs using Nvidia's RTX Spark chip and on workstations running the DGX Spark. Nvidia says such machines, aimed at consumers, programmers and enterprise knowledge workers, are reaching the market now through partners including Microsoft and Dell. The "A9B" in the model name suggests an architecture where only a fraction of total parameters activate per token, a common way to keep large models fast enough for local hardware; that reading comes from the naming convention rather than anything Nvidia stated.
Kari Briski, Nvidia's vice president of Open Models and Software, framed the release around deployment flexibility in the announcement blog post, saying organizations that want to run agents on Nvidia infrastructure now have an open, optimized model for their tier without choosing between capability and deployability.
Nvidia also claims Super VL completes tasks faster than competing models from Anthropic and OpenAI, naming Opus 5, GPT 6.1 Sol, Luna and Astra, when post-trained on specific agentic tasks. Those tasks include document intelligence, financial accounting, scientific experimentation, medical insurance claims and security attack resilience. Vendor benchmark claims of this kind are routine at model launches and usually rest on tuned configurations, so independent replication is the thing to watch. Nvidia did attach a list of companies it says are using Super VL: Skydance, EY, Ramp, Crowdstrike, Perplexity and Salesforce. Named enterprise users are more concrete evidence than a speed chart, though the details of how each deploys the model are not spelled out.
These are not Nvidia's first open-weight models. The company already has a Nemotron family spanning multimodal AI, reasoning, coding, math and enterprise automation. The strategic context is that the big proprietary labs, which are simultaneously Nvidia customers and investments, are increasingly looking at other silicon, including AWS's Trainium, Cerebras and in-house chips. Open models that enterprises post-train on their own data and run in their own data centers give Nvidia another reason for those buyers to keep buying its GPUs. Briski's post made the ownership argument directly: unlike closed frontier models, organizations can post-train Super VL on their own data, deploy it on their own infrastructure and own what they build.
The Bolt models serve a different purpose: seeding demand for AI PCs built on Nvidia chips. The pitch is privacy. Consumers may hesitate to hand cloud agents access to family, health and financial information, so local models let PC users delegate everyday projects without sending those workloads to the cloud, per Briski. In practice, a 74-billion-parameter model running on a single local machine will face memory and power constraints that cloud deployments do not, which is likely why Nvidia offers the much smaller E4B alongside it. Whether buyers treat local agents as a real category or a nice-to-have is an open commercial question, and one Nvidia has a direct interest in answering affirmatively.
Availability: Super VL and Bolt 74B-A9B are out now. Bolt E4B is due later this month, according to Nvidia.
Why it matters: Enterprises that want agents on their own infrastructure gain another open option, and the named early users give the release more weight than a benchmark claim alone. Developers should treat the speed comparisons against Anthropic and OpenAI as vendor-reported until tested on their own tasks. For product teams building on-device features, the Bolt models and RTX Spark machines are the first concrete Nvidia-backed path to private local agents, though the hardware footprint and real-world latency remain to be seen.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Nvidia is leaning hard into the trend of enterprise companies wanting to run their own open models in their own data centers and also trying to make on-device, private AI agents a thing.