NVIDIA has announced a major release of open AI models in the Nemotron 3.5 family, developed specifically for autonomous, around-the-clock agents. The lineup is spearheaded by the Nemotron 3.5 Super VL multimodal neural network with 120 billion parameters. According to the manufacturer, the model delivers comparable accuracy at one-fifth the size of leading industry giants while providing up to a fivefold increase in execution speed. This release fits into NVIDIA's broader strategy to transform its own graphics accelerators and software platform into the foundational infrastructure for autonomous digital agents.
Super VL Flagship: Giant Accuracy in a Compact Size
The Nemotron 3.5 Super VL model is optimized for document, table, and chart recognition tasks, as well as multi-step business processes where an agent must sequentially execute chains of actions without constant human intervention. The release's key positioning is the combination of quality comparable to the largest closed models with significantly higher inference speeds and a lower computational footprint, making the model practical for deployment in enterprise scenarios and on local workstations.
Compact Bolt Models for Local Deployment
Alongside the flagship, two compact local models in the Bolt series have been introduced, designed to run on mobile devices, personal computers, and NVIDIA DGX Spark stations. Their appearance underscores the company's bet on the on-device inference segment: agents can run directly on the end device, which reduces cloud dependence and network latency while simplifying data privacy concerns.
Open Weights, Data, and Fine-Tuning Ecosystem
The models in this family come with open weights, datasets, and fine-tuning files via the NVIDIA NeMo platform. Distribution is handled through Hugging Face and in the format of NVIDIA NIM microservices, allowing developers to quickly integrate the models into their own products. According to the release, major companies such as Salesforce, Atlassian, CrowdStrike, Perplexity, and Dassault Systèmes are already integrating Nemotron 3.5.
Context: Nemotron as Part of NVIDIA's Agentic Roadmap
The 3.5 release continues a line initiated earlier in 2026: at GTC Taipei, the company announced Nemotron 3 Ultra, an open Mixture-of-Experts language model with a total of 550 billion parameters, also focused on autonomous agents and claimed by NVIDIA to be five times faster and 30% cheaper in target scenarios. Its adoption involved Perplexity, Palantir, ServiceNow, and other companies. Thus, the 120-billion-parameter Super VL complements the lineup with a lighter multimodal model for a wide range of tasks, while Ultra covers heavy computational scenarios.
What This Means for the Market
The release of open-weight Nemotron 3.5 strengthens competition in the agentic AI segment and provides companies with an alternative to closed cloud APIs: the ability to fine-tune on their own data and deploy on local GPUs. For the ecosystem, this means an increase in the number of production-ready models capable of independently planning actions, using tools, and carrying multi-step tasks through to completion—a key milestone in the next stage of generative AI evolution beyond classical chatbots.
