Is DGX Spark 64GB enough for the AI work you want to keep local?
A smaller memory configuration creates a more useful buying question than a faster-chip headline: does it fit the work you actually want to keep on your desk? NVIDIA’s October 2 announcement introduces DGX Spark with 64GB of unified memory, scheduled for availability from six manufacturer partners on October 23, starting at $4,999. As of October 8, that is an announced purchase option, with availability still ahead.
The new configuration retains the GB10 Grace Blackwell Superchip, DGX OS and NVIDIA AI software stack used by the 128GB model. NVIDIA says the 64GB version supports models with up to 100 billion parameters entirely on device. The useful distinction is therefore capacity within the same platform: a buyer can choose less memory while retaining that chip and software foundation. The announcement does not establish how every model, context length or combination of simultaneous requests fits within the stated ceiling.
That boundary matters because “local AI” describes where computation happens, while a buying decision needs a workload. A team choosing a private document assistant and a team expecting several agents to work concurrently could reach different memory decisions. Those are illustrative requirements, not reported results from this hardware. My recommendation is to define the intended model, working context and simultaneous use before treating the parameter count as sufficient evidence to buy.
NVIDIA also presents a growth path. It says two 64GB units can pool memory to 128GB through their ConnectX-7 networking and NVIDIA Sync Cluster Assistant, expanding support to models of up to 200 billion parameters. In NVIDIA’s own Qwen 3.8 27B test, the paired systems delivered up to 1.7 times the performance of one system. That is a vendor result for a named test, not a general promise that doubling the machines doubles useful output.
The purchasing implication is straightforward: consider the second unit as another hardware purchase and another capacity option. The announcement provides a starting price for the new configuration, but it does not demonstrate that a pair is the best value against a single 128GB system or a cloud service. Memory expansion and price efficiency are separate claims.
DGX Spark 64GB is most interesting when the intended workload fits one unit and on-device operation has clear value. If the plan already assumes a pair, evaluate that full capacity choice from the outset. The attractive starting point is useful only when it matches the work, rather than deferring a larger purchase that was predictable all along.