NVIDIA’s DGX Spark 64GB systems are scheduled to start reaching customers through partners on Friday, October 23, giving developers another option for running AI models locally. RTX Spark owners can already download Underdog, a desktop agent app that NVIDIA says processes models on the PC without sending user data elsewhere. NVIDIA also plans to introduce NVIDIA Sync Model Launcher later this month to help users install models on individual DGX Spark machines or clusters.
The expansion follows last week’s RTX Spark laptop launch, which brought Windows machines starting at $2,399 with up to 128GB of unified memory. Those laptops paired a Grace CPU and Blackwell RTX GPU, with NVIDIA claiming up to 1 petaflop of AI performance.
Read more: NVIDIA Readies RTX Spark Software for Friday Laptop Launch
The incoming DGX Spark 64GB configuration can run models such as Qwen3.8 27B and Nemotron 3.5 Bolt 74B 24GB, according to NVIDIA. Developers can now manually configure clusters containing up to eight DGX Spark machines, with 1TB of unified memory available for larger models and simultaneous agent sessions. NVIDIA positions those clusters as shared computing resources for small development teams, allowing several developers to work against one model concurrently.
Model Launcher will select and download a model and container appropriate for the hardware configuration, then connect it to OpenCode, NVIDIA says. It will also load DGX Spark Skills to provide maintenance and troubleshooting guidance. The October software update adds Wi-Fi Assistant, which lets owners change networks while their DGX Spark operates in appliance mode.
Larger teams can buy Linux-based DGX Station systems now, while the Windows version is expected from manufacturing partners in December. DGX Station provides 748GB of coherent memory and AI performance rated at up to 20 petaflops. NVIDIA says the Windows machines will support engineering software from companies including Siemens and MathWorks, alongside agents using tools such as Hermes Agent.
Recursive Superintelligence also released Recursive Flint today, a coding model with 40 billion parameters and 2.5 billion active parameters. NVIDIA reports generation speeds exceeding 220 tokens per second on a 64GB RTX Spark. Its weights are downloadable from Hugging Face under Apache 2.0, with support for DGX Spark, RTX Spark and GeForce RTX GPUs.
