NVIDIA @ GTC Berlin 2026: Big upgrades to DGX Spark + Station and RTX Spark - Tech Critter
NVIDIA has announced significant upgrades to its DGX Spark and RTX Spark platforms at GTC Berlin 2026, expanding local AI capabilities across personal, developer, and enterprise tiers.
At GTC Berlin 2026, NVIDIA is showcasing its complete local AI ecosystem, detailing how its three-tier system addresses different user needs. The RTX Spark platform runs personal agents on Windows PCs, DGX Spark supports developers building and scaling apps on Linux, while DGX Station handles enterprise-level workloads, with a Windows version scheduled for December release.
The recently launched RTX Spark is now globally available, designed for personal agents, accelerated content creation, and 1440p gaming on laptops. A new software addition, Underdog, provides a desktop agent app that runs AI models entirely on-device, ensuring user data remains private. Many RTX Spark laptops also feature unlocked bootloaders for developers and enthusiasts to customize their local agents.
DGX Spark Expansion
The 64GB version of DGX Spark machines begins rolling out from partners on October 23, capable of running models like Nemotron 3.5 Bolt 74B 24GB, Qwen3.8 27B, and Gemma 4 26B-A4B. Clustering capabilities have been significantly enhanced, with DGX Spark now supporting up to 8 units in manual configuration, pooling 1TB of unified memory to run frontier-class models like GLM 5.3.
Additional updates include the NVIDIA Sync Model Launcher arriving later this month, which simplifies model and container setup. An October software update will introduce Wi-Fi Assistant for network switching in appliance mode, plus updated DGX Spark Skills for troubleshooting. The hardware will be available through Acer, ASUS, Dell, GIGABYTE, HP, Lenovo, MSI, and NVIDIA.
Enterprise and Model Support
DGX Station offers 748GB of coherent memory and up to 20 petaflops, sufficient to run models like GLM 5.3 and serve 30+ users at cloud-comparable speeds. The latest update extends NVIDIA Sync support to link two Linux-based DGX Station systems. NVIDIA is also announcing new model support, including Recursive Flint, a local coding model with 40B parameters that runs efficiently on RTX hardware.
Day-zero support will be provided for multiple new models including Nemotron Bolt E4B 4GB (released October 23), GLM, DeepSeek, Qwen, and others. Black Forest Labs' FLUX.3 Klein, a 5.6B-parameter image model, can generate 1-megapixel images in about four seconds on an RTX 5090 with just 8GB VRAM. Performance optimizations include faster inference speeds through llama.cpp and vLLM implementations across different hardware configurations.
Related
