Nvidia released open-weight models for enterprise data centers and local AI agents.
Why it matters: The models target businesses that want to customize and run AI on infrastructure they control, rather than rely entirely on public model APIs. The approach could intensify enterprise AI competition while reinforcing demand for Nvidia hardware and software.
- Nemotron 3.5 Super VL has 120 billion total parameters and activates 12 billion per token.
- Nemotron 3.5 Bolt models target local systems, including Nvidia RTX and DGX Spark platforms.
- Nvidia says Super VL can be post-trained on an organization's data, tools, policies and workflows.
- Bolt E4B is scheduled for release on October 23, 2026; Super VL and Bolt 74B-A9B are available immediately.
Nvidia released Nemotron 3.5 Super VL, a 120-billion-parameter open-weight multimodal model for specialized, always-on enterprise agents. The company announced the model on October 15, 2026, alongside two smaller models: Nemotron 3.5 Bolt 74B-A9B 24GB and Nemotron 3.5 Bolt E4B 4GB. Nvidia said Super VL is available immediately, as is Bolt 74B-A9B, while Bolt E4B is scheduled for October 23, 2026.
Super VL is designed for enterprise servers, data centers, workstations and cloud environments. It combines reasoning, instruction following, tool use and visual understanding across text, images, documents and video. Nvidia says organizations can post-train it on their own data, tools, policies and workflows, then deploy it on infrastructure they control. The company says the model was trained on approximately 26 trillion tokens and post-trained with roughly 3.8 million samples.
The Bolt models extend the strategy to local computing. Bolt 74B-A9B is aimed at multistep reasoning, tool calling and web browsing on local Nvidia RTX and DGX Spark systems. Bolt E4B is designed for fast conversational, physical and embedded agents.
Nvidia is releasing the models' weights, datasets and training recipes through its own channels, Hugging Face and related developer platforms. Production deployments can also involve proprietary Nvidia software, including NIM and NeMo. Nvidia identified use cases spanning document intelligence, accounting, scientific experimentation, medical insurance claims, cybersecurity, coding and web research. TechCrunch reported that companies including Scale AI, EY, Ramp and CrowdStrike are users, integrators or customizers.
Independent testing shows both potential and limits. Plotly found Super VL answered 30 of 43 data-analysis questions, with a 12-second median response time, but missed all planted data-quality and hidden-pattern problems in its test. The model's 4-bit version requires about 70GB of memory, according to Plotly.
By the numbers
- 120 billion - Super VL's total parameters
- 12 billion - Parameters activated per token through its mixture-of-experts architecture
- 30 of 43 - Data-analysis questions Super VL answered in Plotly's independent test
Yes, but: The models are open-weight, but enterprise deployment may still depend on proprietary Nvidia software, hardware and support. Independent testing also found reliability gaps, and Nvidia's performance comparisons have not been independently reproduced across identical conditions.
What's next: Nvidia plans to release Nemotron 3.5 Bolt E4B on October 23, 2026.