NVIDIA launches dgx spark 64gb: a high-performance AI desktop for local agents and inference

NVIDIA has introduced the DGX Spark 64GB, a desktop AI system powered by the GB10 Grace Blackwell superchip, designed for local AI agents, fine-tuning, and inference. The system features 64GB of unified LPDDR5x memory, offering enough capacity for today’s most capable 30–35B class open models. It retains the same GPU, CUDA stack, and ConnectX-7 networking as the original DGX Spark, but with reduced memory. NVIDIA also continues to offer the 128GB version for larger single-box workloads, with clustering options for increased memory and compute. The DGX Spark 64GB is equipped with a 20-core Grace Arm CPU and 5th-generation Tensor Cores, delivering up to 1 petaFLOP of FP4 AI compute. Unified memory via NVLink-C2C allows the CPU and GPU to share a single pool, improving efficiency for agents by eliminating the need to copy weights between system RAM and VRAM. The system runs on a standard wall outlet, eliminating the need for server rooms or special cooling. NVIDIA highlights the benefits of running models locally, avoiding per-token fees on cloud APIs. The DGX Spark 64GB supports clustering for 128GB of memory and more compute, with NVIDIA Sync Cluster Assistant simplifying setup. The system is ideal for running agents around the clock, with built-in networking for clustering and support for models like Muse Glimmer and Nemotron 3.5 Lightning.