Summary
NVIDIA says its DGX Spark will be available in a 64GB unified-memory configuration from six manufacturer partners, starting at $4,999 on Oct. 23. Two units can be clustered to provide 128GB of memory for larger local AI workloads.
NVIDIA says a 64GB unified-memory configuration of its DGX Spark personal AI computer will be available from Acer, ASUS, Dell, Gigabyte, HP and MSI starting Friday, Oct. 23, at $4,999. The new option keeps the GB10 Grace Blackwell Superchip, DGX OS and NVIDIA AI software stack used in the 128GB model, while offering less memory for developers running AI workloads locally.
NVIDIA says the system can run models with up to 100 billion parameters on device. That capacity depends on the model and workload: model weights share memory with other runtime needs, such as processing context and concurrent tasks. Local execution can avoid relying on a cloud instance for every job, though it does not by itself guarantee how data is handled by applications or services connected to the system.
How two systems scale
The 64GB model includes NVIDIA ConnectX-7 networking. Two units can be connected directly with a QSFP cable and configured using NVIDIA Sync Cluster Assistant. The company says the cluster pools their memory to 128GB and supports models of up to 200 billion parameters. The assistant detects the connected systems, checks their configuration and sets up the network; NVIDIA says both nodes use the same software stack, so the environment does not need to be reconfigured when moving from one system to two.
NVIDIA reports that two clustered 64GB systems delivered up to 1.7 times the performance of one system in its Qwen 3.8 27B test. It also says the two-unit setup provides twice the memory bandwidth. The announcement does not describe the test methodology, so the performance figure should be read as a result for that company-reported test rather than a general measure across models and tasks.
Software and availability
NVIDIA lists agent development, inference, fine-tuning, data science and edge development among the platform’s intended uses. It says DGX Spark supports NVIDIA Agent Toolkit, CUDA-X AI libraries, Nemotron open models, and frameworks and runtimes including Ollama, vLLM and PyTorch with CUDA. The announcement also names llama.cpp and LM Studio as supported inference-framework downloads.
NVIDIA says a Sync Model Launcher is due at the end of October. It is intended to help launch Qwen3.8 27B on a single DGX Spark or a cluster, and make the model accessible from users’ laptops. The company says the launcher will also set up OpenCode for browser-based coding workflows.
The 64GB configuration will be sold through the six named manufacturer partners at a starting price of $4,999. NVIDIA describes the product as ready for local AI development from setup, while its headline scaling benefit is the option to connect a second unit when a project needs more memory or compute.