NVIDIA DGX Spark Now Supports 8-Way Clusters as Local AI Inference Gets Major Speed Boost
NVIDIA is expanding its local AI ecosystem with major upgrades for DGX Spark, DGX Station, RTX Spark, and RTX PRO systems. The latest updates focus on three areas developers care about most: more memory, faster AI inference, and easier local model deployment.
One of the biggest changes is support for 8-way DGX Spark clusters. Until now, DGX Spark configurations supported up to four systems working together. With the new manual cluster setup, developers can scale to eight systems, opening the door to much larger local AI workloads.
For 64 GB DGX Spark systems, an 8-way cluster can deliver up to 512 GB of pooled memory. For 128 GB DGX Spark systems, the total rises to 1 TB of unified memory. That extra memory is especially important for running larger AI models, longer context windows, multiple AI agents at the same time, and distributed AI development environments.
NVIDIA says the 1 TB memory configuration enables DGX Spark clusters to run frontier-class models such as GLM 5.3. This is a major step for developers and research teams that want to work with advanced AI models locally instead of relying entirely on cloud infrastructure.
The timing is also important because the 64 GB DGX Spark configuration is set to arrive soon, giving more developers a lower-cost entry point into NVIDIA’s local AI hardware ecosystem.
NVIDIA is also introducing new AI models optimized for DGX Spark and RTX systems. One highlight is Nemotron 3.5 Bolt 74B 24GB, which is designed to run well on 64 GB DGX Spark configurations. Other supported models include Qwen3.8 27B and Gemma 4 26B-A4B, giving developers more options for local AI assistants, coding tools, agent workflows, and content generation.
A new NVIDIA Sync Model Launcher is scheduled to arrive later this month. The tool is designed to make local AI model deployment much easier, whether users are working on a single DGX Spark system or a larger cluster. With just a few clicks, Model Launcher can download the right model and container for the available hardware, then connect it to OpenCode.
The launcher will also automatically load DGX Spark Skills, which provide guidance for troubleshooting, device maintenance, workflow help, and access to official documentation inside NVIDIA Sync through OpenCode. This should make DGX Spark more approachable for developers who want powerful local AI without spending extra time manually configuring every component.
The October DGX Spark software update adds more everyday usability improvements. A new Wi-Fi Assistant allows users to switch networks while DGX Spark is running in appliance mode. NVIDIA is also expanding DGX Spark Skills to provide more workflow-focused support for troubleshooting and system maintenance.
RTX Spark is getting attention as well. Recursive Flint, a new local coding model optimized for RTX Spark, is being introduced for developers who want fast on-device code generation. Recursive Flint is a 40B mixture-of-experts model with 2.5 billion active parameters. On 64 GB RTX Spark systems, NVIDIA says it can deliver inference performance of more than 220 tokens per second.
NVIDIA is also adding day-zero support for several newer models across RTX and DGX platforms. These include Nemotron Bolt E4B 4B, GLM, DeepSeek, Qwen, MiniMax M3.1, and FLUX.3 Klein. The supported workloads range from local AI agents and coding assistants to image generation and image editing.
Another major update is the release of new model checkpoints optimized for RAM efficiency on local AI hardware. These checkpoints are designed to let large open models run while preserving enough memory for KV cache, which is essential for longer context lengths and more responsive AI applications.
Each checkpoint is quantized to reduce memory requirements while staying close to full-precision accuracy. Initial examples include Qwen3.8-Flash-Next-FP4-Local, designed for a single DGX Spark 128GB system or RTX PRO 6000, and GLM-5.3-Flash-FP4-Local, designed to run across two systems.
Performance is also improving across NVIDIA’s local AI lineup thanks to optimization work with open-source communities around vLLM and llama.cpp. These improvements target agentic AI workloads, where models often need to reason, call tools, manage context, and respond quickly.
According to NVIDIA, DGX Spark now delivers up to 30% faster performance in Qwen 3.8 27B. DGX Station gains up to a 20% performance boost in GLM 5.3 Flash. RTX PRO 6000 Blackwell sees an even larger improvement, with up to a 40% uplift in Qwen 3.8 Flash-Next.
NVIDIA also shared several broader local AI announcements at GTC 2026 Berlin. Underdog Desktop Agent is coming to RTX Spark, allowing personal AI agents to run entirely on the device so user data stays on the PC. Recursive Flint is bringing open-weight coding capabilities to NVIDIA local AI hardware. New model support is expanding across Nemotron, GLM, DeepSeek, Qwen, MiniMax, and FLUX.3 for local agents, coding, image generation, and editing.
DGX Spark 64GB is scheduled to arrive on Friday, October 23, giving developers another option for building and running local AI applications. DGX Spark will also scale to eight systems, with NVIDIA Sync expected to add eight-system support and Model Launcher integration to simplify model downloads and connection to coding tools.
Enterprise teams are getting new management integrations for DGX Station, including support for tools from Absolute Security, Splunk by Cisco, Fleet Device Management, Microsoft Intune, Perforce Puppet, Progress Software Chef, and Tanium. These integrations are aimed at helping IT departments deploy, monitor, and manage local AI infrastructure more efficiently.
DGX Station for Windows is planned for launch in December. NVIDIA says this will help bring local AI into manufacturing, design, simulation, validation, and engineering automation workflows. The company is also introducing three open-source robotics workflows for DGX Station, focused on refining robot models, training dexterous manipulation skills, and post-training world models for synthetic data generation.
NVIDIA Cloud Functions is also coming to DGX Station in early access. This will allow teams to share local model endpoints through a common interface and update underlying models without changing application code.
Overall, NVIDIA’s latest DGX Spark and RTX platform updates show a clear push toward powerful local AI development. With 8-way DGX Spark clusters, up to 1 TB of unified memory, faster inference, optimized model checkpoints, and easier deployment tools, NVIDIA is making it more practical for developers, researchers, and enterprises to run advanced AI workloads directly on their own hardware.
