What does NVIDIA's RTX Spark change for a PC that sits on a desk?

A new category of Windows computer is arriving in October. NVIDIA calls it RTX Spark.

The machine puts a 1 petaflop Blackwell GPU, a 20‑core Grace CPU, and up to 128 gigabytes of unified memory into a thin laptop or a small desktop. Lenovo showed the Yoga Pro 9n and Yoga 9n 2‑in‑1. Acer showed a compact desktop concept. Six OEMs total will ship in October.

Local agents get a simpler start

Three widely used agent apps now offer one‑click local model setup on Windows. Hermes Agent, OpenClaw, and Perplexity Portable Computer each bundle llama.cpp with NVIDIA inference optimizations. The agent detects the GPU, picks a suitable model, and runs it without manual downloads or quantization tuning.

Hermes Agent works on RTX PCs, RTX PRO workstations, and DGX Spark. OpenClaw requires an RTX GPU with at least 24 gigabytes of VRAM. Perplexity Portable Computer runs on Linux today with 24 gigabytes VRAM; Windows support is coming. All three keep data on the device unless the user explicitly allows cloud escalation.

Inference speed jumps on existing hardware

llama.cpp throughput rises up to 1.9 times on a GeForce RTX 5090. The gain comes from kernel optimizations and enhanced speculative decoding. Faster prefill adds further improvement. vLLM delivers 1.2 times on an RTX PRO 6000 Blackwell Workstation Edition and up to 1.4 times on two DGX Spark clusters. XQA attention kernels in FlashInfer and backend optimizations drive the vLLM gains. Both backends are available directly and through LM Studio and Ollama.

Idle PCs become a shared compute pool

More than half of U.S. households have two or more PCs. Much of that silicon sits idle. NVIDIA Personal AI Router (PAIR) is a free, open‑source tool that discovers compatible machines on a local network and routes independent inference jobs to whichever system has capacity. It works with Ollama and LM Studio. Devices can join or leave the network without breaking the pipeline.

Supported hardware includes GeForce RTX 20 series GPUs and newer. RTX PRO workstation GPUs from Turing architecture onward are included. DGX Spark and Apple M4 or newer silicon also work. A user can ask Hermes to build a “Sunday Reset” plan from a cluttered inbox. Hermes splits the work into subagents. PAIR distributes those subagents across available PCs instead of queuing them on a single GPU.

Creative apps run diffusion models locally

CyberLink PhotoDirector 365 adds an AI PC Mode optimized for RTX Spark. Users get generative editing and image enhancement. The mode also handles object and distraction removal. Background removal and replacement, portrait refinement, and creation of entirely new visuals round out the toolkit. On NVIDIA GPUs PhotoDirector uses TensorRT‑RTX and FP8 to accelerate local AI. The user chooses local or cloud processing per task.

New models sized for local deployment

Nemotron 3.5 Lightning is a 30‑billion‑parameter model that runs on RTX PCs, RTX PRO workstations, DGX Spark, and Jetson. Z.ai’s GLM‑5.3‑Flash brings multimodal mixture‑of‑experts capability to DGX Station. Qwen released Qwen3.8‑Flash‑Next and Qwen3.8‑27B for local agentic and coding workloads on DGX Spark and DGX Station. LTX 2.5 generates open‑world video optimized for RTX GPUs, DGX Spark, and DGX Station with NVFP4, FastVideo, and ComfyUI enhancements. MiniMax‑H3 adds synchronized audio; FastH3 distills it to four steps for a 7 times speedup. Meta’s Muse Glimmer targets coding and agentic tasks on GeForce RTX PCs, DGX Spark, DGX Station, and Jetson. DeepSeek v4 Flash is a 284‑billion‑parameter MoE with 13 billion active parameters that runs on a two‑DGX‑Spark cluster or DGX Station. NVFP4 quantization with DGX Spark support reduces memory pressure for all these models.

What the announcement does not cover

Pricing for RTX Spark systems is not disclosed. Battery life claims for thin laptops are not quantified. The exact release date in October is not specified. How many existing RTX 20‑series owners will adopt PAIR is not projected.