NVIDIA Unveils RTX Spark and DGX Local AI Systems at GTC Berlin
At GTC Berlin, NVIDIA presented its expanding local AI hardware and software lineup, positioning personal PCs, developer systems and enterprise workstations as platforms for private, on-device AI agents. The company’s RTX Spark, launched on Windows PCs on Oct. 16, is designed for personal agents, accelerated content creation and 1440p gaming with up to 128GB of unified memory and 1 petaflop of AI compute. NVIDIA highlighted early reviewer support and said many RTX Spark laptops include unlocked bootloaders to let developers customize local agents. The new Underdog desktop agent app, powered by the Husky inference engine, runs models entirely on-device, and RTX Spark laptops are available from ASUS, Dell Technologies, HP, Lenovo, Microsoft and MSI, with Acer planned. For developers, NVIDIA introduced DGX Spark 64GB systems starting to roll out from partners on Oct. 23. They can run models such as Nemotron 3.5 Bolt 74B, Qwen3.8 and Gemma 4, and can be linked into clusters of up to eight systems through NVIDIA ConnectX-7 networking and Sync Cluster Assistant. A manual cluster delivers up to 1TB of unified memory for advanced coding models, longer contexts and multiple agents. A later Sync Model Launcher will simplify model downloads, container setup and OpenCode connections, while an October software update adds a Wi-Fi Assistant and maintenance guidance through DGX Spark Skills. NVIDIA also positioned DGX Station for larger teams, offering 748GB of coherent memory and up to 20 petaflops of AI performance, enough to serve more than 30 users on systems from ASUS, Dell Technologies, GIGABYTE, HP, MSI, Supermicro and Exxact. The company added enterprise management integrations, open-source robotics workflows, and a coming NVIDIA Cloud Functions early-access feature that lets teams register local model endpoints in a shared catalog. DGX Station for Windows, expected in December, will extend local AI models and agents into professional enterprise applications from vendors including Siemens, Dassault Systèmes, MathWorks and Rockwell Automation. In software, NVIDIA announced Recursive Flint, a 40-billion-parameter mixture-of-experts local coding model with 2.5 billion active parameters. The model delivers more than 220 tokens per second on a 64GB RTX Spark and is available on Hugging Face under Apache 2.0. NVIDIA also expanded day-zero support for Nemotron Bolt E4B, GLM, DeepSeek, Qwen, MiniMax and Liquid AI models, along with Black Forest Labs’ FLUX 3 Image Klein, a 5.6-billion-parameter image model optimized for consumer GPUs. Additional inference optimizations in llama.cpp and vLLM are expected to accelerate agentic tasks on RTX PRO, DGX Spark and DGX Station systems.
