Nvidia is adding a 64GB version of DGX Spark, its compact system for running and developing AI models locally. Acer, Asus, Dell, Gigabyte, HP and MSI will sell the configuration this month with DGX OS and Nvidia’s AI software stack preinstalled.
The machine retains the GB10 Grace Blackwell chip and ConnectX-7 networking found in the 128GB version. Nvidia says one 64GB unit supports models with as many as 100 billion parameters. Two units can connect over a QSFP cable, combine 128GB of memory and support models up to 200 billion parameters. In Nvidia’s Qwen 3.8 27B test, the paired setup reached up to 1.7 times the performance of one system; that result is vendor testing, not a general guarantee across workloads.
Nvidia Sync Cluster Assistant detects the connected machines, checks their configuration and prepares the network. A Model Launcher due at the end of October will initially automate downloading and running Qwen 3.8 27B across one or two systems. The platform also supports Ollama, vLLM, PyTorch, Nvidia’s Agent Toolkit and CUDA-X libraries.
The new option gives developers a lower-memory entry point for private, on-device experiments while preserving a path to a two-node setup. Nvidia did not provide a universal performance claim for other models or applications.