Nvidia’s desk-side AI lineup fills in, from RTX Spark laptops to DGX Station

Nvidia used GTC Berlin to lay out its local AI lineup: RTX Spark laptops for personal agents on Windows, a new 64GB DGX Spark shipping from partners on October 23 (DGX Spark 64GB), clustering of up to eight DGX Sparks, and DGX Station for teams, with a Windows version due in December. Software included a Model Launcher in Nvidia Sync, seven enterprise management integrations for DGX Station, robotics reference workflows and day-zero support for many open models. Nvidia says RTX Spark launched October 16 with up to 128GB of unified memory and 1 petaflop of AI compute (RTX Spark); DGX Station is quoted at 748GB of coherent memory and up to 20 petaflops, and an eight-unit DGX Spark cluster at 1TB of unified memory. Laptops from ASUS, Dell, HP, Lenovo, Microsoft and MSI are listed, with Acer coming.

The significance is less any single machine than a stack for running capable agents on hardware an organization owns, ranging from a 4GB voice model to a desk-side machine Nvidia says can serve more than 30 users.

Local inference changes who holds the data. Underdog, a desktop agent app Nvidia highlighted, says it runs models entirely on device so personal data never leaves the PC. For lawyers, clinicians, engineers handling export-controlled designs or anyone unwilling to route an inbox through someone else’s servers, that is a substantive benefit if apps hold to their claims. Enterprise integrations with tools such as Intune, Splunk and Tanium may matter more than raw speed, because they make desk-side AI look like any other managed endpoint.

Recursive’s Flint, an Apache 2.0 open-weights coding model with 40 billion parameters and 2.5 billion active, featured prominently. Nvidia claims over 220 tokens per second on a 64GB RTX Spark, while Recursive separately reported about 131 tokens per second on an M4 Pro MacBook Pro (Recursive’s Flint); its maker says automatically generated kernels sped generation by 55% without changing weights. All are vendor figures awaiting independent benchmarks.

Every performance figure here comes from Nvidia’s own blog at its own conference. And the software layer, from Sync to Model Launcher to Cloud Functions on DGX Station, is Nvidia’s, so owning the hardware does not mean owning the stack. There is a security dimension too: many RTX Spark laptops ship with unlocked bootloaders to encourage tinkering, and agents on a desk with access to local files and tools need careful permissioning. Moving agents closer to sensitive data reduces some risks and creates others.

If the approach scales, the market could split, with frontier models in the cloud and a growing share of everyday agent work, such as coding, drafting and robotics training loops, on machines in the office. That could pressure per-token cloud pricing for routine tasks and is a healthier balance than total cloud concentration, provided open weights stay open. More in our Nvidia coverage, enterprise coverage and AI agents coverage.