NVIDIA DGX Spark Now Supports Up to 8-Device Clusters with 1TB Unified Memory
At GTC 2026, NVIDIA announced that DGX Spark now supports manual configuration for clustering up to 8 devices, enabling a combined unified memory pool of up to 1TB for the 128GB version. Additionally, software updates deliver up to 40% performance gains for RTX PRO 6000 Blackwell hardware in specific AI workloads. This expansion significantly boosts local AI processing capabilities for enterprise and advanced user workloads by bridging the gap between desktop hardware and cluster-level scalability. By optimizing integration with open-source ecosystems like vLLM and llama.cpp, NVIDIA makes running large-scale local AI models much more practical. The 64GB version of DGX Spark is priced at $4,999 and ships starting October 23, while the new NVIDIA Sync Model Launcher and updated local AI platforms facilitate easier container deployment and automated maintenance. Furthermore, new quantized model checkpoints like Qwen3.8-Flash-Next-FP4-Local allow users to run massive models directly on single or dual systems.
## BACKGROUND
NVIDIA DGX Spark is a compact, ARM-based computing system and platform developed for local artificial intelligence and advanced developer workloads. Frameworks like vLLM and llama.cpp serve as popular open-source engines designed to accelerate large language model inference and serving efficiently across diverse hardware architectures.