NVIDIA Launches 64GB DGX Spark for Local AI at $4,999

NVIDIA has expanded its DGX Spark personal AI supercomputer lineup with a new 64GB configuration, targeting developers and researchers who require capable local inference and agent development without the capital outlay or memory footprint of higher-end models. Shipping globally starting October 23, the new SKU retails from authorized manufacturing partners including Acer, ASUS, Dell, Gigabyte, HP, and MSI at a starting price of $4,999. This release strategically addresses current silicon supply constraints and shifting model optimization trends by providing a streamlined entry point into unified memory local AI systems. Built around NVIDIA’s GB10 Grace Blackwell Superchip, the 64GB DGX Spark maintains the platform’s core architecture while optimizing cost and power efficiency. The system integrates dedicated ConnectX-7 networking, DGX OS, and a comprehensive CUDA-accelerated AI software stack, enabling out-of-the-box compatibility with industry-standard runtimes such as Ollama, vLLM, and PyTorch. By retaining the full software ecosystem of the 128GB variant, NVIDIA ensures that developers can deploy and fine-tune models containing up to 100 billion parameters entirely on-premises, eliminating cloud dependency for initial development and privacy-sensitive workloads. A defining advantage of the DGX Spark platform remains its horizontal scalability. Utilizing the NVIDIA Sync Cluster Assistant, two 64GB units can interconnect via a single QSFP cable to pool memory into a 128GB unified workspace. This clustered configuration supports models up to 200 billion parameters while delivering twice the memory bandwidth and up to 1.7 times the inference throughput compared to a single node. The Sync software automatically validates hardware configurations and provisions the high-speed network, allowing workflows to scale seamlessly without manual infrastructure reconfiguration. An upcoming NVIDIA Sync Model Launcher will further simplify deployment, enabling one-click execution of open models like Qwen3.8 27B across single or clustered nodes. The 64GB launch positions NVIDIA to capture a broader segment of the local AI market, including independent researchers, small teams, and developers who previously required 128GB configurations to run modern open-weight models. This move aligns with industry trends toward parameter-efficient architectures and on-device AI agents, reducing reliance on centralized cloud GPU clusters. Early ecosystem integrations already include Blender, which is preparing a dedicated installer for the platform, alongside comprehensive development playbooks hosted on NVIDIA’s build portal. By introducing a more accessible memory tier without compromising software readiness or cluster scalability, NVIDIA reinforces its strategy to democratize local AI development. The 64GB DGX Spark offers a production-grade alternative to consumer workstations, providing enterprise-level networking, unified memory management, and optimized AI frameworks in a compact desktop form factor. As open models continue to mature and on-device inference becomes standard practice, this configuration establishes a flexible foundation for scalable, privacy-focused artificial intelligence workflows.