Nvidia Unveils New Open-Weight Models for Enterprise and Local AI Agents

Image: Nemotron 3.5 Super VL delivers frontier-level intelligence in its class for high-volume agentic workflows. · Source

Nvidia introduces Nemotron 3.5 Super VL and Bolt models, open-weight AI agents designed for enterprise data centers and local PCs, promising faster task completion and privacy-focused deployments.

Nvidia has announced the release of new open-weight AI models aimed at enabling always-on intelligent agents across enterprise servers and local edge devices. The flagship Nemotron 3.5 Super VL is a 120-billion-parameter model optimized for enterprises, offering a balance of efficiency and accuracy in instruction-following and tool-use tasks. Despite having fewer parameters than Nvidia's larger Nemotron 3 Ultra model, Super VL delivers comparable performance in domains such as document intelligence, financial accounting, scientific experimentation, medical claims processing, and cybersecurity resilience.

Complementing Super VL, Nvidia introduced two more compact Bolt models designed for on-device AI agents capable of running locally on Nvidia-powered PCs and workstations equipped with the RTX Spark chip and DGX Spark systems. The Bolt 74B-A9B 24GB model strikes a balance between speed and reasoning for tasks like multistep reasoning, tool invocation, and web browsing, while the smaller Bolt E4B 4GB model is tailored for conversational agents. These models aim to support privacy-conscious users and organizations by enabling AI workloads to be processed without sending data to the cloud.

Nvidia emphasizes the models' open nature, allowing enterprises to post-train Super VL on proprietary data and deploy the models within their own infrastructure, providing greater control, security, and privacy than closed, proprietary AI models. This strategy aligns with a broader industry trend favoring customizable AI deployments in enterprise data centers.

Performance claims state that Super VL completes agentic tasks up to five times faster than competing models from Anthropic and OpenAI, with Nvidia citing partnerships with industry players such as Skydance, EY, Ramp, CrowdStrike, Perplexity, and Salesforce who have adopted or customized Super VL for specific workflows.

The Nemotron family now includes the recently released Lightning and Ultra models, covering high-throughput and advanced reasoning needs, while the new Super VL and Bolt models extend the lineup to efficient, edge-focused AI.

All Nemotron 3.5 models support multimodal inputs, understanding text combined with visual information like charts, tables, and documents, enabling continuous context retention across complex workflows without needing multiple AI systems. The models are open weight, with training datasets and recipes publicly available, facilitating extensive customization through Nvidia’s NeMo toolkit.

The open models are accessible via platforms such as Hugging Face, Open Router, and ModelScope, and integrate with Nvidia’s hardware stack, including DGX Spark, DGX Station, GeForce RTX systems, and Nvidia Jetson for edge applications. Bolt 74B-A9B is already available, with Bolt E4B scheduled for release later this month.

Nvidia aims to support a diverse spectrum of AI agent deployments, from large-scale enterprise servers to consumer-level PCs, enhancing both capability and deployability without compromising data control or user privacy.