NVIDIA Releases Lower-Memory DGX Spark for AI Developers

NVIDIA has officially unveiled a new 64-gigabyte (GB) memory variant of its DGX Spark desktop AI supercomputer. This release aims to make high-performance local AI development more accessible to a broader range of developers, enabling the creation of large language models and AI agents directly on their workstations.

According to NVIDIA, AI agents represent the fastest-growing segment within the open-source community. The surge in demand for running advanced AI models locally stems from the escalating costs associated with cloud-based services. DGX Spark is designed to address common developer pain points, including insufficient graphics memory, complex cluster deployments, and software compatibility issues.

The new 64GB model utilizes GB10 Grace-Blackwell supercomputing chips and features a unified memory architecture. It supports the smooth deployment of popular open-source models such as Gemma4 26B and Qwen3.8-27B, while providing ample memory for handling long-context inference tasks. NVIDIA also updated the pricing for its existing 128GB version. The 64GB model starts at $4,999 and is scheduled for release on October 23.

DGX Spark offers native multi-node clustering capabilities through its integrated high-speed network card and NVIDIA Sync software suite. This allows developers to set up clusters of up to four DGX Spark units without requiring extensive networking expertise. Performance tests indicate that a two-node cluster of 64GB DGX Sparks can deliver up to a 1.7x increase in inference throughput. The platform is supported by NVIDIA's CUDA acceleration software stack and optimizations for key inference frameworks like vLLM and llama.cpp.

The DGX Spark aims to reduce the cost and complexity of AI innovation by enabling local development. This approach mitigates cloud computing expenses, enhances data privacy, and allows for faster iteration cycles. The introduction of the more affordable 64GB model makes powerful local AI computing capabilities available to a wider audience of researchers and developers.