NVIDIA Launches 64GB DGX Spark for More Affordable Local AI
NVIDIA is expanding its DGX Spark local AI platform with a new 64GB unified memory configuration, giving developers and AI enthusiasts a lower-cost alternative to the existing 128GB model.
Starting at $4,999, the 64GB DGX Spark will launch on October 23 through six OEM partners: Acer, ASUS, Dell, Gigabyte, HP, and MSI. The system retains the same GB10 Grace Blackwell Superchip and core software stack as the higher-capacity configuration, but targets local AI workloads that do not require 128GB of unified memory.
The new configuration reflects a broader shift in local AI hardware requirements. While large unified memory capacities remain important for running and fine-tuning larger models, advances in model efficiency are making lower-memory systems increasingly practical for inference, agentic workloads, development, and other targeted use cases.
🖥️ 64GB DGX Spark Hardware Configuration #
The 64GB DGX Spark is built around NVIDIA’s GB10 Grace Blackwell Superchip, with its core hardware architecture largely unchanged from the 128GB configuration.
Key specifications include:
- 64GB unified CPU-GPU memory
- 20-core Arm CPU
- 273 GB/s shared CPU-GPU memory bandwidth
- ConnectX-7 RDMA network interface
- DGX OS
- Full NVIDIA AI software stack
NVIDIA states that the platform can run models with up to 100 billion parameters on-device, depending on model architecture, quantization, context requirements, and workload characteristics.
The system is designed for workloads including local inference, AI agents, fine-tuning, data science, and edge AI development.
The primary difference from the existing DGX Spark configuration is therefore memory capacity rather than a fundamental change to the compute platform.
💰 Lowering the Entry Barrier for Local AI #
The introduction of the 64GB model reflects changing requirements across local AI workloads.
Large language models have traditionally driven demand for systems with substantial memory capacity because model weights, KV caches, intermediate activations, and other runtime data can consume significant amounts of memory. However, not every local AI workload requires the same memory footprint.
For example, dense models such as Qwen 3.8 27B have been demonstrated running within 32GB of memory, although available memory can still constrain context capacity and other runtime requirements.
This creates a gap between entry-level local AI hardware and systems designed specifically for very large models. The 64GB DGX Spark is positioned to address that middle ground.
Recent memory pricing fluctuations also make high-capacity configurations more expensive. Offering a lower-memory variant allows NVIDIA and its OEM partners to reduce the initial hardware commitment for users whose workloads do not justify 128GB.
💵 Pricing Compared With 128GB Systems #
The 64GB DGX Spark starts at $4,999, making it substantially less expensive than many currently available 128GB-configured GB10 systems.
Depending on memory and storage availability, supply and demand, and inventory conditions, 128GB systems have generally been selling in the $7,000–$9,000 range.
The price difference gives developers more flexibility to match hardware capacity to actual workload requirements rather than paying for memory that may remain unused.
For users primarily performing inference, development, agent experimentation, or smaller-scale fine-tuning, 64GB can therefore represent a more targeted entry point into NVIDIA’s local AI ecosystem.
🔗 Multi-Node Expansion Remains Available #
One of the key advantages of the 64GB DGX Spark is that reducing local memory capacity does not eliminate its ability to scale across multiple systems.
Because the platform retains the ConnectX-7 RDMA NIC, multiple DGX Spark systems can be connected into a high-speed cluster.
This provides an alternative to purchasing a single system with maximum memory capacity from the beginning. Users can start with a 64GB configuration and add additional nodes if their model size, memory requirements, or inference throughput increases.
A multi-node configuration can therefore provide additional aggregate memory and compute capacity while allowing the initial hardware investment to remain relatively low.
🛠️ NVIDIA Adds Cluster Assistant #
NVIDIA is also simplifying the process of connecting multiple DGX Spark systems.
The company has added Cluster Assistant to its NVIDIA Sync software. The feature can automate foundational ConnectX-7 network configuration and SSH setup across multiple DGX Spark devices.
Previously, configuring a multi-node DGX Spark cluster could require users to manually manage network topology, SSH configuration, scripts, and terminal commands. Automating these steps reduces the amount of infrastructure work required before the cluster can be used.
However, Cluster Assistant should not be confused with a complete workload orchestration system.
The feature handles the underlying network and connectivity setup but does not automatically deploy inference or fine-tuning workloads. Once the cluster is configured, users still need to use frameworks and workflows such as NCCL, vLLM, and PyTorch to run distributed workloads.
The current Cluster Assistant implementation supports configurations ranging from two to four DGX Spark systems.
🤖 Model Launcher Expands NVIDIA Sync #
NVIDIA Sync also gains a Model Launcher feature designed to simplify model deployment across one or multiple DGX Spark systems.
The feature can help users download and launch supported models without manually completing every installation and configuration step.
NVIDIA is also integrating the workflow with OpenCode, its browser-based coding agent, to make it easier to establish agentic development environments on DGX Spark systems.
Together, these tools push DGX Spark toward a more integrated local AI development experience, where hardware configuration, model deployment, and AI-assisted development can be managed through a unified workflow.
⚙️ 64GB vs. 128GB Configuration Strategy #
The 64GB DGX Spark does not replace the existing 128GB model. Instead, NVIDIA is positioning the two configurations for different workload requirements.
| Configuration | Primary Positioning | Typical Workload Considerations |
|---|---|---|
| 64GB DGX Spark | Lower-cost local AI platform | Inference, agents, development, data science, smaller models |
| 128GB DGX Spark | High-memory local AI platform | Larger models, higher memory requirements, demanding fine-tuning |
The 128GB configuration remains the more appropriate choice for workloads that require substantial memory headroom, particularly when running larger models or performing memory-intensive fine-tuning.
The 64GB configuration, meanwhile, provides a more accessible option for developers who want the GB10 platform without paying for maximum memory capacity.
📊 A More Granular Local AI Hardware Market #
The 64GB DGX Spark represents a shift toward more differentiated local AI hardware configurations.
Rather than assuming that every AI developer needs the maximum available memory capacity, NVIDIA is offering multiple configurations based on workload size, budget, and scalability requirements.
The ability to combine multiple 64GB systems also changes the purchasing equation. Developers can begin with a smaller configuration and expand to a multi-node setup as their requirements grow, rather than committing to a high-capacity system at the outset.
As local AI workloads continue to diversify, this approach could make dedicated AI development hardware more accessible while preserving a path toward larger distributed deployments.
