Nvidia moves running smart agents to offices with local computers and open programming models

Nvidia disclosed at the GTC conference in Berlin an integrated system of on-premise AI hardware for offices, shifting the workload of running independent agents from central cloud computing centers to developers’ machines and teams directly. The system includes RTX Spark devices aimed at Windows PCs, DGX Spark computers for Linux environments, and DGX Station workstations designed for large enterprises.

This architecture aims to break the exclusive reliance on cloud interfaces.Thus, RTX Spark computers provide unified memory of up to 128 GB and processing power of one petaflop of AI calculations, with a unlocked bootloader that lets developers build and customize local agents freely. Nvidia backed the system with the UnderDog application, which uses a proprietary inference engine to run agents entirely on the device without user data leaving it, while CEO Jensen Huang showcased training of the MicroDoc robot from Haging Face locally using the Isaac Lab library and the Neuron 3.5 Bolt 74 B model.

In the developer and software-team track, DGX Spark systems with 64 GB capacity, slated for release through partners on October 23, allow linking up to eight units together via Connect X 7 networks and an integrated aggregation assistant. This aggregation gives teams a unified memory capacity of a full terabyte, sufficient to run advanced models such as GLM 5.3 and DeepSec V 4 Flash locally, while distributing concurrent work sessions of multiple developers across a single model in parallel. The system also adds tools to launch models and connect them to software development environments with a few clicks, as well as automated maintenance guides and a wireless-network assistant.

Advanced workstations become independent internal inference serversIt offers DGX Station devices with coherent memory of 748 GB and processing power up to 20 petaflops, enabling service for more than thirty users simultaneously at speeds comparable to cloud solutions. Nvidia is preparing a Windows-specific version for December that integrates directly with major engineering and industrial software suites, alongside support for enterprise administrative control tools to manage security compliance, and provides local cloud-like functions that let teams share endpoints compatible with OpenAI interfaces without modifying source code.

The hardware is accompanied by the release of open-source models optimized for on-premise devices, led by the Recursiv Flint programming model comprising 40 billion parameters but activating only 2.5 billion parameters per call, which achieved generation speeds exceeding 220 tokens per second on desktop Spark hardware with advanced results in agent-evaluation benchmarks. The system also provided immediate support for lightweight, compressed models in a four-bit floating-point format, including versions of Neuron, Flux 3 and Quoin, together with software updates to open-source inference engines that boosted agent task efficiency by 20-40 percent.

Our analytical reading sees this development reshaping infrastructure calculations for tech teams and enterprises in the Gulf, Egypt and the Levant; organizations that face regulatory constraints prohibiting the transfer of their sensitive data and logs to external servers are now able to run fully capable AI agents on-premises, reducing recurring cloud-inference bills and protecting operational privacy. It also gives software and independent-system developers in the region the freedom to train and test agents with short cycles, without tying their operational continuity to constant internet connectivity or exclusive reliance on international providers.