NVIDIA announced the DGX Spark 64GB, a new 64GB configuration of its GB10-powered desktop AI system built by partners including Acer, ASUS, Dell, Gigabyte, HP, and MSI. The system delivers up to 1 petaFLOP of FP4 AI compute using a Blackwell GPU paired with a 20-core Grace Arm CPU and unified LPDDR5x memory over NVLink-C2C. Designed to run open models and always-on local AI agents without metered cloud API fees, the hardware runs on a standard wall outlet and includes ConnectX-7 networking to allow clustering two units for 128GB of memory.
The system ships with DGX OS preinstalled with the NVIDIA AI stack, including PyTorch, Jupyter, Ollama, and single-command installation for NVIDIA NemoClaw. The 64GB unified memory architecture avoids copying weights between system RAM and VRAM, letting several models, key-value caches, and tool processes share a single address space. This configuration is positioned to handle 30B to 35B class open models like quantized builds of Meta's Muse Glimmer, enabling developers to run always-on personal agents, perform QLoRA fine-tuning on local codebases, and execute local model evaluation benches.
Why it matters: Providing 1 petaFLOP of local desktop AI compute with unified memory lets developers run continuous agent workflows and fine-tune models entirely on-premise without recurring per-token cloud costs.