The new configuration keeps the GB10 Grace Blackwell Superchip, DGX OS and the full NVIDIA software stack, and NVIDIA says it can run models of up to 100 billion parameters entirely on device. The 20-core Arm CPU complex and 273 GB/s of shared memory bandwidth carry over unchanged from the 128GB model, as does the ConnectX-7 200GbE network card. That networking matters because two 64GB units can be clustered over a QSFP cable to pool 128GB of memory, expand model support to 200 billion parameters and, in NVIDIA's Qwen 3.8 27B test, deliver up to 1.7x the performance of a single system.

To make that clustering less of a chore, NVIDIA is adding a Sync Cluster Assistant that detects connected units, validates their configuration and sets up the ConnectX-7 network automatically. A Sync Model Launcher, due at the end of the month, will download and start models such as Qwen 3.8 27B across one or more systems and hook them up to the OpenCode browser-based coding agent. Runtimes including Ollama, vLLM, llama.cpp and LM Studio are supported out of the box, and a prebuilt Blender installer is promised soon.

The timing is no accident. Tom's Hardware notes that capable dense models such as Qwen 3.8 27B now fit within 32GB of RAM, so 128GB is no longer essential for local inference, while soaring memory prices make permanently pairing the GB10 chip with costly LPDDR5X a barrier to entry. ServeTheHome reports that the 128GB DGX Spark, which launched at $3,999 about a year ago, now sits at roughly $6,950, so the cheaper configuration arrives as the original climbs well past its launch price. ServeTheHome also cautions that $4,999 for 64GB still feels high, and questions whether two 64GB nodes beat a single 128GB system, though NVIDIA's pitch is that clustering needs no software reconfiguration.