Nvidia ships open-weight Nemotron 3.5 Super VL for enterprise agents and Bolt models for local PCs

Nvidia this week released two families of open-weight models: a large one built for always-on agents inside enterprise data centers, and a pair of lightweight models meant to run agents locally on new AI PCs and workstations.

The enterprise model is Nemotron 3.5 Super VL, a 120-billion-parameter system. Nvidia says it uses fewer parameters than its Nemotron 3 Ultra model while delivering similar accuracy on instruction-following and tool-use tasks — a trade-off aimed squarely at agents that stay running and keep calling tools.

At the other end sit two Bolt models. Nemotron 3.5 Bolt 74B-A9B 24GB carries 74 billion parameters and is designed for local agentic work, balancing speed against multistep reasoning, tool calling and web browsing. The smallest, Bolt E4B, targets conversational agents.

Kari Briski, Nvidia's vice president of open models and software, wrote in the announcement blog post that any organization wanting to run agents on Nvidia infrastructure now has an open, optimized model for its tier without having to choose between capability and deployability.

She framed the open weights as a question of control: 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.

Nvidia paired the launch with benchmarking claims and a customer list. It says that after post-training on specific agentic tasks — document intelligence, financial accounting, scientific experimentation, medical insurance claims and security attack resilience — Super VL completes tasks faster than competing models from Anthropic and OpenAI. Skydance, EY, Ramp, CrowdStrike, Perplexity and Salesforce are named as users.

The two tiers map onto two separate market bets. Enterprises increasingly want models they can host and tune themselves, while consumer AI assistants remain constrained by privacy worries; running agents locally, Nvidia argues, lets PC users delegate everyday projects without sending those workloads to the cloud. The hardware is arriving alongside the models, with AI PCs running the RTX Spark chip and DGX Spark workstations shipping through partners including Microsoft and Dell.

For Nvidia the play is as much about silicon as software. As frontier labs spread their compute across AWS Trainium, Cerebras and chips of their own design, open models help anchor where agents actually run — and whose hardware they run on. Super VL and Bolt 74B-A9B are available now, with Bolt E4B due later this month; the open questions are how much post-training customers actually do and whether local agents find buyers beyond developers.