NVIDIA expands Nemotron 3.5 from data centers to local agent systems
NVIDIA has released two Nemotron 3.5 models and scheduled a third for October 23, spanning server, desktop and edge deployments. The company publishes weights, datasets and training recipes, while analysts see both open-ecosystem support and a strategy to strengthen demand for NVIDIA hardware.
In this article
Three models for different deployment tiers
NVIDIA has expanded the Nemotron 3.5 family with models aimed at enterprise servers, local workstations and smaller edge systems. Nemotron 3.5 Super VL is a 120-billion-parameter vision-language model for multimodal agents and long-running workflows. Nemotron 3.5 Bolt 74B-A9B is designed to fit into 24 GB through native two-bit quantization, while Bolt E4B targets conversational and physical agents in a 4 GB footprint.
According to NVIDIA, Super VL and Bolt 74B-A9B are available now, while Bolt E4B is due on October 23. The company is releasing model weights, datasets and training recipes, and distributing the models through services including Hugging Face. NVIDIA presents this as an open, customizable range that can run from data-center infrastructure down to RTX, DGX Spark and Jetson-class devices.
Local agents and vendor-reported performance
NVIDIA says Bolt 74B-A9B is intended for everyday local agent tasks such as research, writing and code debugging without sending the workload to the cloud. The vendor also says Super VL can process text and visual information from webpages, charts, tables and PDFs in one model. It reports that Super VL performs at roughly one-fifth the parameter count of its largest reasoning model and claims throughput up to five times higher than the next model in its class.
Those figures come from NVIDIA’s own launch material and selected benchmarks; they are not independent measurements. The company also reports that Super VL was trained on about 26 trillion tokens, then post-trained with roughly 3.8 million samples and reinforcement learning across multiple environments.
Openness also reinforces NVIDIA’s platform
AI Business cites Omdia analyst Alexander Harrowell, who sees two motives behind the release. Open models give developers more options for customization and help the broader open-source ecosystem compete with lower-cost Chinese open-weight models. At the same time, optimized models can make NVIDIA GPUs, orchestration software and deployment tools more attractive.
That strategic link matters as Google, Amazon and other companies integrate their own accelerators with models and software. NVIDIA recently open-sourced Run.ai and is extending Nemotron across its hardware stack. Harrowell also notes a practical obstacle: enterprises must become comfortable again with packaged AI software running on local PCs or shared devices rather than only through cloud services.
The release therefore combines a product launch and an ecosystem strategy. It gives organizations more control over deployment and adaptation, but “open” does not remove operational work. Teams still need suitable hardware, security controls, model evaluation and a reason to prefer local execution over a managed service.
