NVIDIA DGX Spark 64GB Desktop for Local AI Agents

NVIDIA announced a new 64GB configuration of its DGX Spark desktop AI system on October 23, 2026, designed to empower developers with local execution capabilities for AI models and agents. This updated system, available through Acer, ASUS, Dell, Gigabyte, HP, and MSI, utilizes the GB10 Grace Blackwell superchip and offers a starting point for running local models, with the option to cluster two 64GB units for expanded memory and compute power. The core message from NVIDIA is to enable developers to run open models and always-on agents directly on their desks, bypassing the per-token billing associated with cloud APIs. This move addresses the significant growth in token consumption by AI agents, which has increased 14x since early 2026 due to factors like tool calls, retries, long context windows, and multi-step planning. By providing dedicated hardware, NVIDIA aims to offer a more cost-effective solution for continuous agent operation.

The DGX Spark 64GB model retains the GB10 Grace Blackwell superchip, the NVIDIA CUDA accelerated AI software stack, and ConnectX-7 networking found in the original system. The key difference is the reduction to 64GB of unified LPDDR5x memory, down from the 128GB in the previous configuration. NVIDIA positions this 64GB capacity as sufficient for running today's most capable open models in the 30–35 billion parameter class. For workloads requiring more memory or compute within a single unit, NVIDIA continues to offer the 128GB DGX Spark. The ability to cluster multiple 64GB units allows users to scale their memory and processing power incrementally as their needs grow.

Internally, the GB10 Grace Blackwell superchip integrates a Blackwell GPU with 5th-generation Tensor Cores and a 20-core Arm CPU. This configuration delivers up to 1 petaFLOP of FP4 AI compute performance, factoring in sparsity. The system specifications for the DGX Spark 64GB include a 20-core Arm CPU (composed of 10× Cortex-X925 and 10× Cortex-A725 cores), up to 1 petaFLOP FP4 AI compute with sparsity, 64GB of coherent unified LPDDR5x memory with 273 GB/s bandwidth, and storage options of 1, 2, or 4TB of self-encrypting NVMe M.2 drives. Networking is handled by a ConnectX-7 NIC offering 200GbE speeds, alongside Wi-Fi 7 and Bluetooth 5.3. It features one HDMI 2.1a display output and runs on the NVIDIA DGX OS, which is Ubuntu-based. The compact form factor measures 150 × 150 × 50.5 mm and weighs 1.2 kg, supporting local models up to 100 billion parameters. The unified memory architecture, where the CPU and GPU share a single memory pool over NVIDIA's NVLink interconnect, is a critical design element enabling efficient data access and processing for demanding AI tasks.