NVIDIA Adds a 64GB DGX Spark at $4,999 for Local AI

NVIDIA has announced a 64GB unified-memory configuration of DGX Spark, expanding the desktop Grace Blackwell platform below the existing 128GB model. Partner systems from Acer, ASUS, Dell, Gigabyte, HP and MSI are scheduled to become available on October 23, 2026, starting at $4,999.

The smaller-memory configuration retains the GB10 Grace Blackwell Superchip, DGX OS, NVIDIA AI software stack and ConnectX-7 networking. The main tradeoff is memory capacity: NVIDIA says the 64GB system supports models up to 100 billion parameters on device, while larger workloads may require a second system or a higher-memory configuration.

NVIDIA is positioning clustering as part of the smaller configuration's upgrade path. Two DGX Spark 64GB systems can connect through their ConnectX-7 interfaces, pooling memory to 128GB and doubling aggregate memory bandwidth. NVIDIA says this expands supported model size to as much as 200 billion parameters.

In NVIDIA's own Qwen 3.8 27B test, two clustered 64GB systems delivered up to 1.7x the performance of one system. That figure is a vendor measurement for a specific workload, not an independent benchmark or a guarantee of 1.7x scaling across models, runtimes or inference settings.

The accompanying NVIDIA Sync Cluster Assistant detects connected DGX Spark systems, validates their configuration and sets up the ConnectX-7 network. NVIDIA says each node retains the same software environment when scaling from one unit to two.

A separate Sync Model Launcher is planned for the end of October. NVIDIA says it will download and launch Qwen 3.8 27B on a single Spark or cluster and configure OpenCode to use the locally hosted model. That feature is not yet generally available at the time of this announcement.

The new configuration is not a faster DGX Spark. Its value proposition is a lower entry price for developers whose models and context fit inside 64GB while preserving the same GB10 compute and networking foundation.

Memory remains a hard capacity boundary for local AI. Model weights are only part of the footprint; runtime overhead, KV cache, context length and concurrent workloads also consume memory. NVIDIA's parameter-count limits therefore should not be interpreted as a promise that every model below a given parameter count will fit under every precision, context and runtime configuration.