NVIDIA RTX Spark Launches Globally, Designed for Local AI Agents

NVIDIA officially launched the RTX Spark globally on October 16 at 8 PM Beijing time. The chip integrates Blackwell RTX graphics, CUDA ecosystem and AI capabilities into a single package, with up to 128GB of unified memory, targeting on-device personal agents and large language models.

The first wave of RTX Spark laptops comes from Lenovo, ASUS, MSI, Dell and Microsoft, split into two tiers: high-end models with 6144-core GPUs and mainstream models with 5120-core GPUs. The high-end tier offers 32GB to 128GB memory options, while the mainstream tier, mostly from Dell and Microsoft, starts at 24GB and 32GB. The ASUS ProArt Creation 16 RTX Spark 128GB edition, priced at 37,999 yuan, sold out quickly after launch.

RTX Spark combines a Blackwell GPU with a Grace CPU, co-designed with MediaTek, and lets the CPU and GPU share a memory pool of up to 128GB. Under FP4 precision, it delivers up to 1 Petaflop of AI compute. NVIDIA says the platform can run personal agents with up to 125 billion parameters locally, and offline LLMs with 120 billion parameters. Game performance is claimed to be up to 9x higher than common Windows SoCs, with AI image generation up to 11x faster.

Software support arrived quickly: Adobe is adding TensorRT acceleration to AI features in Photoshop, Premiere and Substance 3D, while CapCut Pro, Bilibili Live and D5 Lite are supported on day one. Agents such as GitHub Copilot, Perplexity and Pi offer guided setup. RTX Spark laptops start at 18,999 yuan. Jensen Huang said the RTX Spark brings NVIDIA AI, RTX graphics and CUDA to Windows PCs, enabling users to create, game and run personal AI agents.