NVIDIA Expands Local AI Hardware Lineup at GTC Berlin With New Software and Clustering

NVIDIA announced significant updates to its local AI hardware portfolio at GTC Berlin, focusing on expanded clustering capabilities and new software integrations. The company is emphasizing on-device intelligence across its RTX Spark, DGX Spark, and DGX Station product lines to serve consumers, developers, and enterprise teams.

The DGX Spark system now supports clusters of up to eight units, providing 1TB of unified memory for running frontier-class models like GLM 5.3. This update allows small development teams to share a single cluster backend while running concurrent sessions against the same model, according to NVIDIA blog posts.

Software and Model Updates

NVIDIA Sync Model Launcher will become available later this month to simplify access to local AI on DGX Spark systems. The tool automatically downloads suitable models and containers, connecting them to OpenCode while loading DGX Spark Skills for troubleshooting and documentation. This aims to reduce the complexity of setting up local inference environments for developers.

The October software update for DGX Spark adds Wi-Fi Assistant capabilities and updates DGX Spark Skills. Day-zero support is provided for new Nemotron, GLM, DeepSeek, Qwen, MiniMax, and FLUX.3 AI models. These updates include new accelerations for llama.cpp and vLLM inference engines, along with a collection of checkpoints optimized for local hardware performance.

Consumer and Enterprise Hardware

RTX Spark, which launched on October 16, powers personal agents on Windows PCs with up to 128GB of unified memory. The Underdog desktop agent app is now available for RTX Spark, running models entirely on-device via its Husky inference engine to keep user data private. Many RTX Spark laptops feature an unlocked bootloader to help developers customize local agents.

DGX Station improves manageability with seven new integrations and three new open-source reference workflows for robotics. NVIDIA Cloud Functions is coming in early access to DGX Station. The Windows version of DGX Station launches in December, supporting many manufacturing applications and agent harnesses for enterprise teams.

What should you do?

Security and technology professionals should review the new clustering capabilities of DGX Spark to assess their impact on local AI development workflows. Teams using RTX Spark should evaluate the privacy implications of running personal agents locally with the Underdog application. Organizations planning enterprise deployments should monitor the December release of DGX Station for Windows for manufacturing and agent harness support.

What to do and how to stay safe: NVIDIA

• Evaluate whether local AI clustering with DGX Spark fits your team's model size and concurrency requirements.
• Review the day-zero model support list to determine if your current inference engines benefit from the new accelerations.
• Check if your RTX Spark devices have unlocked bootloaders if you plan to customize local agent environments.
• Monitor the December release date for DGX Station on Windows to plan enterprise manufacturing workload migrations.

Step-by-step guide: Mobile Device Security: How to Protect Data Beyond Passwords

General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.

Frequently asked questions

How many DGX Spark units can be clustered together?

DGX Spark now supports clusters of up to eight units in manual configuration, providing 1TB of unified memory for larger models.

When is DGX Station for Windows available?

The Windows version of DGX Station is scheduled to launch in December, supporting manufacturing apps and agent harnesses.

What new AI models have day-zero support?

Day-zero support includes Nemotron, GLM, DeepSeek, Qwen, MiniMax, and FLUX.3 AI models for NVIDIA local hardware.