NVIDIA has introduced a 64GB configuration of the DGX Spark desktop AI system, built around the GB10 Grace Blackwell superchip. This version keeps the same 1 petaFLOP FP4 AI compute, 20‑core Grace Arm CPU, and ConnectX‑7 200GbE networking as the original 128GB model, but uses 64GB of unified LPDDR5x memory instead of 128GB. Unified memory means the CPU and GPU share a single pool over NVLink‑C2C, eliminating costly data copies and giving agents a single address space for models, KV caches, and tool processes.
The 64GB DGX Spark is sized for today’s most capable 30‑35B class open models such as Qwen 3.8‑27B and Nemotron 3.5 Lightning. Because the system runs on a standard wall outlet and needs no special cooling, developers can keep an always‑on agent on their desk without incurring per‑token API charges. Token consumption for agents has grown 14x since early 2026, making local ownership a cost‑effective alternative to metered cloud services.
When workloads expand, two 64GB units can be clustered over ConnectX‑7 using NVIDIA Sync, delivering 128GB of pooled memory and up to 1.7x the performance of a single 128GB Spark. The clustering setup is native — no extra hardware or switches are required for two‑node configurations, and the Sync assistant simplifies network and SSH management.
Practical uses on a single box include: an always‑on personal agent that triages issues and summarizes papers, QLoRA fine‑tuning of a coding model on a private repo, day‑one evaluation of newly released open weights, multi‑model agent teams that share unified memory, and edge‑robotics prototyping where a vision transformer is validated locally before deployment to Jetson devices.
The 64GB DGX Spark is available October 23 through NVIDIA Marketplace, OEM partners, and retail, while the 128GB version remains offered for larger single‑box needs.
#AI #Product #NVIDIA #DGXSpark #Agents #EdgeComputing