NVIDIA introduced a 120-billion-parameter enterprise model and two smaller Bolt models, with two offerings available now and one expected later in October.
NVIDIA has introduced new open-weight models aimed at enterprise agents and AI workloads running directly on computers. The lineup includes Nemotron 3.5 Super VL, a 120-billion-parameter model for enterprise use, and two smaller models in the Bolt family for local deployment. The announcement was published October 19, 2026.
NVIDIA says Super VL can deliver accuracy similar to its Nemotron 3 Ultra on instruction-following and tool-use tasks despite having fewer parameters. The company also claims that, after post-training for specific agent tasks, the model completes work faster than named competing models from Anthropic and OpenAI. Those tasks include document intelligence, financial accounting, scientific experimentation, medical insurance claims and security attack resilience. NVIDIA named Skydance, EY, Ramp, CrowdStrike, Perplexity and Salesforce among organizations using Super VL.
For local computing, Nemotron 3.5 Bolt 74B-A9B 24GB is a 74-billion-parameter model designed for multistep reasoning, tool use and web browsing. The smaller Bolt E4B is intended for conversational agents. NVIDIA says the models can run on AI PCs equipped with its RTX Spark chip and on workstations using DGX Spark. The article notes that partner systems from companies including Microsoft and Dell are entering the market.
The local models are designed to let users delegate tasks without sending those workloads to cloud services. For organizations, NVIDIA says Super VL can be post-trained on company data and deployed on their own infrastructure. That approach gives teams more control over where workloads run, though the article does not provide independent benchmark validation of NVIDIA’s comparative performance claims.
NVIDIA said Super VL and Bolt 74B-A9B are available now; Bolt E4B was scheduled to arrive later in October. The releases add to NVIDIA’s broader Nemotron family, which includes models for areas such as multimodal AI, reasoning, coding and enterprise automation.