NVIDIA Brings Its Local AI Line to GTC Berlin: RTX Spark, DGX Spark and DGX Station

NVIDIA used GTC Berlin to present its full lineup of local AI hardware, splitting the portfolio by where the work happens: RTX Spark for personal agents on Windows PCs, DGX Spark for developers building on Linux, and DGX Station for enterprise teams running frontier workloads.

RTX Spark, released on October 16, is the consumer-facing piece. NVIDIA describes up to 128GB of unified memory, one petaflop of AI compute and 1440p gaming at 100 frames per second, with laptops from ASUS, Dell, HP, Lenovo, Microsoft and MSI available and Acer coming soon. Many of those machines ship with an unlocked bootloader so developers can customise local agents.

On the software side, Underdog is now available for RTX Spark. The desktop agent runs on its Husky model-specific inference engine and executes entirely on device, so user data never leaves the PC. During the keynote Jensen Huang demonstrated agents on RTX Spark training Hugging Face's Microduck robot using Isaac Lab, Hermes Agent and Nemotron 3.5 Bolt 74B.

DGX Spark gets a 64GB configuration rolling out from partners on October 23, capable of running models including Nemotron 3.5 Bolt 74B 24GB, Qwen3.8 27B and Gemma 4 26B-A4B. Built-in ConnectX-7 networking and NVIDIA Sync Cluster Assistant let developers link up to eight units in manual configuration, reaching 1TB of unified memory, enough for frontier-class models such as GLM 5.3.

Management software is filling in alongside the hardware. NVIDIA Sync's Model Launcher arrives later this month, downloading a matching model and container and wiring it to OpenCode, while the October DGX Spark update adds a Wi-Fi Assistant for appliance mode and DGX Spark Skills for troubleshooting and maintenance guidance.

DGX Station covers the enterprise end, with seven new integration and manageability additions, three new open-source robotics reference workflows, and early access to NVIDIA Cloud Functions. A Windows version is scheduled for December, with support for a range of manufacturing applications and agent harnesses.

The model layer is moving in step. Recursive Flint, a new local coding model, is available now, and NVIDIA claims day-0 support for new Nemotron, GLM, DeepSeek, Qwen, MiniMax and FLUX.3 releases, alongside accelerations for llama.cpp and vLLM and a collection of checkpoints tuned for local hardware.

The bigger shift is that local AI is moving from a single box on a desk to clusters that teams can manage, share and point their agents at. The open questions are how quickly the Windows DGX Station ships and whether Cloud Functions turns a developer workstation line into something enterprises can operate at team scale.