Nvidia launches 64GB DGX Spark for local AI development

A new entry point for local AI

Nvidia officially released the 64GB DGX Spark on October 23, offering a more compact option for developers and IT teams. This new configuration provides a way to run AI models locally without the high cost of the original 128GB system. The starting price for this model is $4,999, though final costs may vary based on the manufacturer and regional pricing.

The hardware is available through six different manufacturing partners, including Acer, ASUS, Dell, Gigabyte, HP, and MSI. Because there is no single universal specification sheet, buyers must confirm specific details like storage, ports, and warranty terms with their chosen vendor. This variety allows organizations to choose a hardware partner that fits their existing infrastructure.

Despite the smaller memory capacity, the system retains the core architecture of the original DGX Spark. It uses the same GB10 Grace Blackwell Superchip and the full Nvidia software platform. This means users still get access to the DGX OS and a suite of tools designed for local inference and model experimentation.

Memory trade-offs and capabilities

The primary difference between the new model and the 128GB version is the unified memory capacity. Nvidia states that the 64GB model can support AI models with up to 100 billion parameters. In contrast, the larger 128GB configuration was advertised as supporting models with up to 200 billion parameters.

While these numbers represent vendor-stated capabilities, they are not guarantees that every model of that size will perform well. Factors such as quantization, context length, and runtime overhead all influence how much memory a specific workload actually requires. The larger 128GB version remains the better choice for memory-intensive tasks like fine-tuning or running very long contexts.

The 64GB version is positioned as a more practical choice for teams whose primary models fit within a smaller memory pool. It provides a starting point for running coding agents and experimenting with models like Nemotron. Developers can use established tools such as Ollama, vLLM, and PyTorch with CUDA to manage their local workloads.

Clustering and networking

Nvidia designed the DGX Spark to be scalable for users who eventually need more memory. Two 64GB systems can be connected directly using built-in ConnectX-7 networking ports and a QSFP cable. When linked, these systems provide 128GB of aggregate unified memory, allowing them to support models with up to 200 billion parameters.

To simplify this process, Nvidia provides the Sync Cluster Assistant. This tool helps the software detect connected systems, verify their configurations, and set up the necessary network. In Nvidia's reported tests using the Qwen 3.8 27B model, a pair of clustered 64GB systems delivered up to 1.7 times the performance of a single unit.

However, buyers should note that a two-unit cluster is not identical to a single computer with 128GB of unified memory. Distributed workloads require specific software and communication between machines, which can introduce overhead. For tasks that do not require shared memory, Nvidia's PAIR tool can distribute independent jobs across multiple computers instead.

Buying decisions for IT teams

Choosing between the 64GB and 128GB models requires a careful look at specific workload requirements. IT teams should verify the actual memory requirements of their intended models, including the specific format and context window. Relying solely on parameter counts may not provide a complete picture of how a model will run in a production environment.

Performance testing is also critical before committing to a large rollout. Teams should test representative prompts and response times to see how the system handles concurrent requests. Additionally, it is important to confirm the manufacturer's specific SSD capacity and available ports to ensure the hardware meets the team's storage needs.

Finally, the total cost of ownership must be considered when deciding between one large system and a cluster of smaller ones. A pair of 64GB systems costs at least $9,998 before adding cables, power, and administration costs. Nvidia plans to release a Sync Model Launcher later this month to help simplify the process of downloading and running models on these systems.