NVIDIA Expands DGX Spark to 8-Way Clusters with 1TB Memory, Boosts AI Inference by 40%

NVIDIA Scales DGX Spark to 8-Way Clusters

NVIDIA announced at GTC Berlin that DGX Spark platforms now support 8-way clustered configurations, doubling the previous 4-way limit2

2

. This expansion allows 64GB DGX Spark systems to pool 512GB of memory, while 128GB configurations can reach an impressive 1TB of unified memory1

1

. The increased capacity enables developers to run frontier-class models like GLM 5.3, support longer contexts, deploy concurrent agents, and facilitate distributed development across small teams sharing a single cluster as a backend2

2

.

40% AI Inference Performance Boost Across Systems

NVIDIA delivered significant AI inference performance improvements across its local AI hardware systems through optimizations with the vLLM and llama.cpp open-source communities2

2

. DGX Spark now achieves 30% faster performance running Qwen 3.8 27B, while DGX Station sees a 20% boost with GLM 5.3 Flash2

2

. The RTX PRO 6000 Blackwell delivers the most dramatic improvement with a 40% uplift in Qwen 3.8 Flash-Next performance, accelerating agentic workflows and large-scale AI workloads2

2

.

New 64GB DGX Spark Systems and Model Support

The 64GB DGX Spark systems begin rolling out from partners on Friday, October 23, providing developers another entry point for local AI development1

1

. These systems can run models including Nemotron 3.5 Bolt 74B 24GB, Qwen3.8 27B, and Gemma 4 26B-A4B1

1

. NVIDIA introduced day-zero support for new AI models spanning Nemotron, GLM, DeepSeek, Qwen, MiniMax M3.1, and FLUX.3 Klein across RTX and DGX systems2

2

. The company also released a collection of model checkpoints optimized for RAM efficiency, including Qwen3.8-Flash-Next-FP4-Local for single DGX Spark 128GB systems and GLM-5.3-Flash-FP4-Local for two-system configurations2

2

.

NVIDIA Sync Updates Simplify Cluster Management

NVIDIA Sync will add support for eight DGX Spark systems and introduce Model Launcher later this month1

1

. Model Launcher simplifies access to local AI on single DGX Spark systems or clustered configurations by downloading models and containers suited to the setup with just a few clicks, then connecting them to OpenCode1

1

. The October DGX Spark software update introduces a Wi-Fi Assistant enabling users to switch networks when running in appliance mode1

1

. Updated DGX Spark Skills provide workflow-focused guidance for troubleshooting, device maintenance, and access to official documentation directly through NVIDIA Sync via OpenCode2

2

.

RTX Spark Brings Personal Agents to Windows PCs

RTX Spark, which launched on October 16, runs personal agents on Windows PCs with up to 128GB of unified memory and 1 petaflop of AI compute1

1

. Underdog, a new desktop agent app built for fully local, private AI, now supports RTX Spark through its Husky model-specific inference engine, running models entirely on device so personal data never leaves the PC1

1

. Recursive Flint, a new local coding model optimized for RTX Spark, offers fast inference performance exceeding 220 tokens per second on 64GB RTX Spark systems2

2

. During the GTC Berlin keynote, Jensen Huang demonstrated how AI agents running on RTX Spark can train Hugging Face's Microduck robot using NVIDIA Isaac Lab, Hermes Agent, and Nemotron 3.5 Bolt 74B1

1

.

Enterprise Management and DGX Station Updates

NVIDIA introduced seven new enterprise management integrations for DGX Station, including Absolute Security, Splunk by Cisco, Fleet Device Management, Microsoft Intune, Perforce Puppet, Progress Software Chef, and Tanium2

2

. These integrations help IT teams deploy and manage local AI infrastructure more effectively. DGX Station for Windows launches in December with support for manufacturing applications and agent harnesses, bringing local AI into design, simulation, and validation workflows2

2

. DGX Station offers 748GB of coherent memory and up to 20 petaflops of AI performance, enabling developers to run advanced models like GLM 5.3 and serve 30-plus users at speeds comparable to cloud models1

1

. These updates position NVIDIA's local AI hardware systems as viable alternatives to cloud-based solutions for organizations prioritizing data privacy, reduced latency, and control over their AI infrastructure.