Nvidia launched the Nemotron 3.5 open-weight models on Thursday, introducing the 120-billion-parameter Super VL for enterprise agents alongside the lightweight Bolt models for local AI PCs. The Super VL model delivers high accuracy in instruction-following and tool-use tasks while completing assignments faster than competing models from Anthropic and OpenAI when post-trained on specific agentic workflows. These workflows cover areas like document intelligence, financial accounting, scientific experimentation, medical insurance claims, and security attack resilience, with early adoption reported by companies such as Skydance, EY, Ramp, Crowdstrike, Perplexity, and Salesforce.
At the same time, the company released two lightweight Bolt models intended to run locally on workstations and new AI PCs powered by the Nvidia RTX Spark chip. These local systems, rolled out through hardware partners including Microsoft and Dell, allow privacy-focused consumers and enterprise knowledge workers to delegate everyday projects without sending sensitive workloads to the cloud. The 74-billion-parameter Bolt 74B-A9B is built to balance speed with reasoning for multistep tasks, tool calling, and web browsing, while the smaller Bolt E4B focuses on conversational agents.
Nvidia is leaning into the growing enterprise demand for deployable, open models that organizations can post-train on their own data and run on private infrastructure. By providing both massive enterprise-tier infrastructure models and efficient local PC alternatives, the GPU giant aims to capture diverse agentic workloads across different computing tiers. This launch matters because it gives organizations and consumers flexible, open-weight options to run autonomous agents locally or on-premise without relying on closed c