At GTC Berlin, NVIDIA unveiled a series of enhancements to its on-premises AI ecosystem, focusing on the RTX Spark, DGX Spark, and DGX Station platforms, through new hardware, software optimizations, and model integrations.
The Underdog Desktop Agent application was announced for the RTX Spark platform, designed to run AI models entirely on the local device, avoiding data transfer outside the PC. On the software development front, Recursive Flint, an open-weighted programming model developed by Recursive Superintelligence and optimized for the company's hardware, was introduced. Launch support covers models such as Nemotron, GLM, DeepSeek, Qwen, MiniMax, and FLUX.3 for local agent workflows, code generation, and image production. Workload efficiency also benefits from optimizations in llama.cpp and vLLM, aimed at increasing inference speed and reducing memory requirements for managing large contexts.
The DGX Spark system with 64GB of memory will be shipping starting Friday, October 23. This configuration scales to eight manually coordinated systems, reaching 1TB of unified memory. An NVIDIA Sync update expected later this October will add support for this eight-unit configuration, along with the Model Launcher tool, which makes it easier to download models and connect them to programming tools. The October software update for DGX Spark also introduces a Wi-Fi assistant for switching networks in device mode, as well as system diagnostics and maintenance tools.
Seven enterprise management integrations have been implemented for DGX Station, including Absolute Security, Splunk by Cisco, Fleet Device Management, Microsoft Intune, Perforce Puppet, Progress Software Chef, and Tanium. The Windows version of DGX Station will be available in December for use in design, simulation, validation, and engineering automation workflows. The platform includes three open-source workflows for robotics, focusing on refining control models, learning manipulation skills, and training models for synthetic data generation. Finally, the Early Access program will provide access to NVIDIA Cloud Functions on DGX Station for sharing local model endpoints through a unified interface, without requiring changes to application code.