NVIDIA Expands DGX Spark to 8-Node Clustering, Reaching 1TB Unified Memory

IT Home, October 21 — NVIDIA announced today at the GTC 2026 conference in Berlin that DGX Spark, which previously supported clustering of up to 4 units, now allows manual configuration for clusters of up to 8 units. When configured with the 128GB version, the unified memory capacity can reach up to 1TB.

At the same time, NVIDIA released updates to its local AI platform and announced collaborations with the llama.cpp and vLLM open-source communities to optimize AI inference performance. Under specific workloads, the inference performance of the RTX PRO 6000 Blackwell improved by up to 40%, while the DGX Spark and DGX Station saw performance gains of up to 30% and 20%, respectively.

NVIDIA stated that the 64GB version of DGX Spark, priced at $4,999 (IT Home note: currently about 33,450 RMB based on current exchange rates), will begin shipping through partners on October 23. For cluster deployments, the NVIDIA Sync Model Launcher is scheduled to go live later this month. Through this tool, users can download models and containers matching their hardware configurations and connect the models to OpenCode. The tool will also automatically load DGX Spark Skills to provide guidance on device troubleshooting, maintenance, and workflows.

In terms of software updates, the October update for DGX Spark adds a Wi-Fi Assistant, allowing devices to switch wireless networks while in appliance mode. DGX Spark Skills will also be integrated into NVIDIA Sync via OpenCode to support daily maintenance and troubleshooting.

Regarding models, NVIDIA announced support for more AI models capable of running locally, including Nemotron, GLM, DeepSeek, Qwen, MiniMax, and FLUX.3 Klein, covering tasks such as agents, coding, image generation, and editing.

NVIDIA also introduced Recursive Flint, a coding model optimized for local AI hardware. Featuring a Mixture-of-Experts (MoE) architecture, the model has a total of 40 billion parameters with 2.5 billion active parameters. NVIDIA claims that the model achieves an inference speed of over 220 tokens per second on an RTX Spark equipped with 64GB of memory.

To address memory pressure when running large models locally, NVIDIA also released a set of memory-optimized model checkpoints that reduce memory footprint through quantization while maintaining accuracy as close to full-precision models as possible. Among them, Qwen3.8-Flash-Next-FP4-Local can run on a single 128GB DGX Spark or RTX PRO 6000, while GLM-5.3-Flash-FP4-Local is designed for dual-system configurations.

Inference performance optimization was another key focus of this update. According to data released by NVIDIA, in llama.cpp, speculative decoding optimizations with multi-token prediction (MTP) for Qwen 3.8 Flash Next can speed up end-to-end agentic tasks on the RTX PRO 6000 Blackwell by up to 1.4x. In vLLM, speculative decoding and FlashInfer prefill kernel optimizations for Qwen3.8 27B can accelerate related tasks on DGX Spark by up to 1.3x. DGX Station running GLM 5.3 Flash achieved up to a 1.2x increase in throughput.

NVIDIA also announced enterprise management and application expansion plans for DGX Station, including seven new management integrations such as Absolute Security, Cisco's Splunk, Microsoft Intune, and Tanium, helping enterprises deploy and manage local AI infrastructure.

The Windows version of DGX Station is expected to launch in December, supporting manufacturing-related applications as well as design, simulation, verification, and engineering automation workflows.

Additionally, NVIDIA plans to launch open-source robotics reference workflows for DGX Station and offer NVIDIA Cloud Functions in early access, allowing teams to share local model services through a unified interface.

Related reading: "NVIDIA's Jensen Huang: Over 1,200 Apps Tested, RTX Spark Is the World's Only Computer to Run All Apps Locally and Natively" "A New Choice for Desktop AI Supercomputing: NVIDIA DGX Spark 64GB Memory Edition Officially Released, Priced at $4,999"