Bringing the Spark to Europe: Local Intelligence Powers Next Wave of Agents

NVIDIA unveiled its RTX Spark and DGX Spark hardware line‑up at GTC Berlin, enabling fully local personal agents on Windows PCs and scalable multi‑GPU clusters for developers on Linux. The RTX Spark platform delivers up to 128 GB unified memory, 1 peta‑FLOP AI compute, and supports the Underdog Desktop Agent powered by the Husky inference engine, while DGX Spark 64 GB systems (shipping Oct 23) can run large models such as Nemotron 3.5 Bolt 74B 24 GB, Qwen 3.8 27 B and Gemma 4 26 B‑A4B and scale to eight‑node clusters via NVIDIA Sync. New software updates add Wi‑Fi Assistant, DGX Spark Skills, and Model Launcher, plus accelerations for llama.cpp and vLLM and a collection of checkpoints optimized for local AI hardware. These announcements give engineers a turnkey path to build, run, and manage private, high‑performance AI agents without cloud dependency.

⚡ Key Takeaways

• RTX Spark offers up to 128 GB unified memory and 1 peta‑FLOP AI compute for on‑device agents.
• Underdog Desktop Agent runs fully locally on RTX Spark using the Husky model‑specific inference engine.
• DGX Spark 64 GB supports models like Nemotron 3.5 Bolt 74B 24 GB and can cluster up to eight units via NVIDIA Sync.
• New software adds Wi‑Fi Assistant, DGX Spark Skills, Model Launcher, and accelerations for llama.cpp and vLLM inference engines.
• DGX Station gains seven new integrations and a Windows version launching in December for enterprise workloads.

Engineers can now deploy private, high‑throughput AI agents on commodity PCs or small‑scale GPU clusters, reducing latency and data‑privacy concerns while retaining the ability to run state‑of‑the‑art LLMs locally.

✅ Practical Steps

1. Download and install the Underdog Desktop Agent from the RTX Spark download page to run local agents on a Windows PC.
2. Deploy a DGX Spark 64 GB system (available Oct 23) and use NVIDIA Sync Cluster Assistant to connect up to eight nodes for larger model inference.
3. Enable the new Model Launcher and Wi‑Fi Assistant features via the latest DGX Spark software update to accelerate llama.cpp and vLLM workloads.

Want the full story? Read the original article.