NVIDIA announced local AI hardware updates at GTC Berlin. RTX Spark now runs personal agents on Windows PCs with 128GB memory and 1 petaflop compute. Underdog Desktop Agent is available for RTX Spark. DGX Spark 64GB systems launch on October 23. DGX Spark now supports clusters of up to eight units. NVIDIA Sync adds Model Launcher and Wi-Fi Assistant this month. DGX Station gains seven new IT management integrations. DGX Station for Windows arrives in December with manufacturing app support. Three new open-source robotics workflows are available for DGX Station. NVIDIA Cloud Functions comes to DGX Station in early access. Recursive Flint, a new local coding model, is now available. Day-zero support adds Nemotron, GLM, DeepSeek, Qwen, MiniMax, and FLUX.3 models. New accelerations improve llama.cpp and vLLM inference engines.
What changed
Local AI hardware lacked unified memory, clustering, and Windows support for enterprise workloads.
What it unlocks
Running frontier models locally on PCs, clusters, and workstations with private, high-performance inference.
- 128GB unified memory on RTX Spark
- 1 petaflop AI compute on RTX Spark
- 748GB coherent memory on DGX Station
- 20 petaflops AI performance on DGX Station
What you need to act on it
- RTX Spark, DGX Spark, or DGX Station hardware
- Windows or Linux OS depending on system
- NVIDIA Sync for clustering
Sources