NVIDIA adds a 64GB DGX Spark, with two-system clustering for local AI
NVIDIA announced on October 2, 2026 that DGX Spark will be available in a new 64GB unified-memory configuration. Systems from Acer, ASUS, Dell, Gigabyte, HP and MSI are scheduled to arrive on October 23, starting at $4,999 in the United States.
Up to 100B-parameter models on one system
The 64GB configuration keeps the GB10 Grace Blackwell Superchip, DGX OS and NVIDIA AI software stack used by the larger configuration. NVIDIA says a single unit can run models up to roughly 100 billion parameters locally, including agentic workloads.
Popular local AI runtimes and frameworks including Ollama, vLLM, llama.cpp and PyTorch are supported, making the device a purpose-built platform for local inference, fine-tuning and agent development.
Two systems can form a 128GB cluster
Each DGX Spark includes ConnectX-7 networking. NVIDIA Sync Cluster Assistant can configure two directly connected 64GB units as a 128GB cluster, expanding support to models up to roughly 200 billion parameters.
In NVIDIA’s own Qwen 3.8 27B testing, a two-system configuration reached up to 1.7x the performance of one unit. Real-world scaling will depend on the model and workload.
U.S. availability
The new 64GB systems are scheduled for October 23 in the U.S. through major hardware partners, with pricing starting at $4,999. Availability and exact configurations will vary by manufacturer.
KX3 perspective: local AI becomes more modular
The important shift is not just a lower-memory DGX Spark. NVIDIA is making it easier to begin with one local AI system and add a second node later when models or agent workloads need more memory.
At $4,999, this remains a specialist machine rather than a mainstream PC. Its value will depend on whether developers actually need large unified memory, CUDA tooling and multi-node local inference rather than a conventional RTX workstation or cloud instance.
Announcement and source
Announcement date: October 2, 2026
Primary source: NVIDIA Blog — “NVIDIA DGX Spark 64GB Gives Developers More Ways to Build and Scale Local AI”
Image: Unsplash / Avel Chuklanov
[…] Read this article in English […]