NVIDIA expands local AI hardware line with RTX Spark, DGX updates

NVIDIA used GTC Berlin on Oct. 21, 2026, to broaden its local AI hardware push across consumer, developer and enterprise systems, with new software, model support and partner availability updates for RTX Spark, DGX Spark and DGX Station.

Evidence and context

In a blog post published Wednesday, NVIDIA said RTX Spark, which launched on Oct. 16, is aimed at personal agents and local AI on Windows PCs, while DGX Spark is positioned for developers on Linux and DGX Station for enterprise teams running frontier workloads on Linux. The company said a Windows version of DGX Station is due in December. The announcement also named several specific software and model additions, including Underdog Desktop Agent for RTX Spark, a DGX Spark 64GB rollout starting Friday, Oct. 23, and new day-zero support for models from Nemotron, GLM, DeepSeek, Qwen, MiniMax and FLUX.3.

NVIDIA said RTX Spark systems can include up to 128GB of unified memory and 1 petaflop of AI compute, and that several laptop partners, including ASUS, Dell Technologies, HP, Lenovo, Microsoft and MSI, have systems available now. The company described DGX Spark 64GB systems as support for models such as Nemotron 3.5 Bolt 74B 24GB, Qwen3.8 27B and Gemma 4 26B-A4B, and said clusters of up to eight units are supported in manual configuration. NVIDIA also said DGX Spark can provide 1TB of unified memory across an eight-unit cluster, allowing larger models and concurrent agents.

For larger shared systems, NVIDIA said DGX Station offers 748GB of coherent memory and up to 20 petaflops of AI performance, while a Windows version expected from partners in December will extend that platform into Windows-native enterprise workflows. The company said DGX Station for Windows will support manufacturing software from nTop, Dassault Systèmes, Keysight Technologies, MathWorks, Rockwell Automation, Siemens, Synopsys, Unity and Altera. NVIDIA also said new integrations with Absolute Security, Splunk by Cisco, Fleet Device Management, Microsoft Intune, Perforce Puppet, Progress Software Chef and Tanium are meant to simplify enterprise deployment and device management.

Several of the announcements are tied to later release dates. NVIDIA said DGX Spark 64GB systems will begin rolling out on Oct. 23 from partners, that NVIDIA Sync Model Launcher will arrive later in October, and that DGX Station for Windows is expected in December. NVIDIA also said NVIDIA Cloud Functions will arrive in early access on Linux-based DGX Station, letting teams register local model endpoints in a shared catalog and access them through an OpenAI-compatible entry point. The company did not provide pricing in the material supplied.

The post also highlighted Recursive Flint, a 40-billion-parameter mixture-of-experts coding model released by Recursive Superintelligence and optimized for NVIDIA local hardware. NVIDIA said the model can exceed 220 tokens per second on a 64GB RTX Spark system and that open weights are available on Hugging Face under the Apache 2.0 license. Separately, NVIDIA said new optimizations for llama.cpp and vLLM could improve specific workloads on its local hardware, including up to 1.4 times faster end-to-end agentic tasks on NVIDIA RTX PRO 6000 Blackwell in llama.cpp and up to 1.3 times faster agentic tasks with Qwen3.8 27B on DGX Spark in vLLM.

The announcement underscores NVIDIA's effort to position local AI systems as an alternative to cloud-only development for coding, robotics and enterprise inference. The material supplied is a product launch and roadmap update rather than a financial disclosure, and it does not include revenue, shipment totals or demand figures.