Nvidia Details Local Agent Tools and RTX Spark PCs for IFA 2026

The numbers

Nvidia’s post ties its throughput claims to specific hardware rather than general marketing language. The llama.cpp backend gets up to a 1.9x speedup on a GeForce RTX 5090 through kernel changes, better speculative decoding, and faster prefill. The vLLM backend gets 1.2x on an RTX PRO 6000 Blackwell Workstation Edition card and up to 1.4x across two clustered DGX Spark systems, helped by new XQA attention kernels in FlashInfer. Both backends already ship the optimizations, and the gains carry through to users running models via LM Studio or Ollama rather than requiring a separate download.

The hardware side of the announcement centers on NVIDIA RTX Spark, a compact Windows PC built around what Nvidia calls a 1 Petaflop RTX Blackwell GPU paired with a 20-core Grace CPU and up to 128GB of unified memory. Lenovo and Acer are building machines around it, joining OEMs that committed earlier in the year, and the first units are due in October alongside a Windows Agent framework meant to let agents run in the background under OS-level control.

PAIR itself is free and open source, distributed as a beta for Windows, macOS, and Linux, and works across GeForce RTX 20-series GPUs and newer, RTX PRO workstation cards from the Turing generation on, DGX Spark, and Apple M4 or newer silicon. It only routes jobs between machines already running Ollama or LM Studio, so it adds no new inference engine of its own, just a scheduler that notices when a task’s home GPU is busy and points it at whichever machine on the network has spare capacity.

None of this is priced or licensed separately from the hardware it runs on, so the practical effect for now is limited to owners of recent Nvidia silicon willing to run the beta software. Nvidia’s post does not disclose adoption figures for PAIR or say how many households the router has actually reached since release, leaving open how much of the “idle PC” capacity it describes gets used in practice.