Nvidia Releases Open-Source Model Matrix: 120-Billion-Parameter Super VL Targets Enterprise Agents
Santa Clara, October 19 — Nvidia released a new portfolio of open-source models covering both enterprise and local PC scenarios. According to TechCrunch, the flagship Nemotron 3.5 Super VL has 120 billion parameters and targets enterprise agent scenarios; Nvidia says its accuracy on instruction-following and tool-calling tasks is comparable to the larger Nemotron 3 Ultra. The two lightweight Bolt models run directly on-device: the 74B-parameter Bolt 74B-A9B is suited to RTX Spark AI PCs and DGX Spark workstations, handling multi-step reasoning, tool calling and web browsing; the smallest, Bolt E4B, is positioned as a conversational assistant and will launch later this month. Super VL and Bolt 74B-A9B are already available.
Kari Briski, Nvidia's vice president of open-source models and software, said in a blog post that, after post-training on specific agent tasks, Super VL completes tasks such as document intelligence, financial bookkeeping and scientific experiments faster than comparable models from Anthropic and OpenAI (Opus 5, GPT 6.1 Sol, Luna, Astra). Early-adopting enterprises include Skydance, EY, Ramp, CrowdStrike, Perplexity and Salesforce. Unlike closed frontier models, enterprises can post-train Super VL with their own data and deploy it on their own infrastructure.