NVIDIA Unveils RTX Spark PCs, Faster Local AI and PAIR at IFA 2026
NVIDIA unveils RTX Spark PCs, faster local inference, PAIR networking and simpler Hermes, OpenClaw and Perplexity agents at IFA 2026 ahead of October launches.
Summary
On September 3 at IFA 2026, NVIDIA, Microsoft and partners expanded local AI. Windows setup is available now for Nous Research’s Hermes Agent, which detects RTX or DGX GPUs and configures llama.cpp, with Linux next. OpenClaw’s Windows app targets RTX GPUs with at least 24GB VRAM, while Perplexity Portable Computer runs on Linux at that minimum and adds Windows soon. Hermes serves millions, and OpenClaw has more than 380K GitHub stars. Perplexity bundles models and tools, avoids credits, and seeks permission before using 15+ cloud models. llama.cpp delivers up to 1.9x throughput on GeForce RTX 5090; vLLM reaches 1.2x on RTX PRO 6000 Blackwell Workstation Edition and 1.4x on two DGX Spark clusters, directly or through LM Studio and Ollama. Free, open source NVIDIA PAIR distributes requests across networked PCs; its beta supports Windows, macOS, Linux, GeForce RTX 20 Series or newer, RTX PRO Turing or newer, DGX Spark and Apple M4 or newer.
RTX Spark Windows PCs ship in October from six existing OEMs, with Acer’s compact desktop concept and Lenovo’s Yoga Pro 9n and Yoga 9n 2-in-1 joining. The platform combines a 1 petaflop RTX Blackwell GPU, up to 128GB unified memory, a 20-core Grace CPU and Windows Agent framework; Electronic Arts, Embark and Ubisoft join May supporters KRAFTON, NetEase, Riot Games and XBOX. CyberLink PhotoDirector 365’s AI PC Mode launches alongside it, using TensorRT-RTX and FP8 for local or cloud generative editing. August releases include 30-billion-parameter NVIDIA Nemotron 3.5 Lightning and Meta Muse Glimmer, 27-billion-parameter Qwen3.8-27B, 284-billion-parameter DeepSeek v4 Flash with 13 billion active parameters, Z.ai GLM-5.3-Flash, Qwen3.8-Flash-Next, LTX 2.5 and MiniMax-H3 with synchronized audio. FastH3 improves MiniMax-H3 performance 7x, with optimized RTX and DGX Spark recipes coming soon; NVIDIA also released NVFP4 quantization with DGX Spark support for Muse Glimmer.
Positives
- llama.cpp raises throughput by up to 1.9x on GeForce RTX 5090 through kernel, speculative decoding and prefill improvements.
- NVIDIA PAIR freely pools compatible computers across Windows, macOS and Linux for parallel local inference.
- Perplexity Portable Computer completes workflows without credits and requests permission before sending content to 15+ cloud models.
- RTX Spark combines a 1 petaflop Blackwell GPU, up to 128GB unified memory and a 20-core Grace CPU.
- FastH3 improves MiniMax-H3 video generation performance 7x through four-step distillation.
- OpenClaw has more than 380K GitHub stars, giving NVIDIA and Microsoft’s simplified Windows setup a large potential audience.
Risks & concerns
- OpenClaw and Perplexity require at least 24GB VRAM, leaving lower-memory RTX systems outside their announced support.
- Perplexity Portable Computer’s Windows support remains forthcoming, while Hermes Agent still awaits Linux support.
- RTX Spark PCs and PhotoDirector AI PC Mode will not arrive until October 2026.
- Cloud escalation can transmit Perplexity task content off-device, although the application requests permission first.
- PAIR remains a beta and supports only specified NVIDIA GPU generations, DGX Spark and Apple M4 or newer silicon.