NVIDIA DGX Spark and Station Bring Open-Source AI Models to the Desktop with Petaflop Performance

NVIDIA has announced at CES that its DGX Spark and DGX Station deskside supercomputers are now capable of running some of the largest open-source and frontier AI models directly from a local system. This marks a significant step forward in making advanced AI development more accessible, particularly for developers who need to work with large-scale models without the overhead of data center infrastructure.

Key Technical Changes and Why They Matter

• Local Desktop Supercomputing: DGX Spark can handle 100-billion-parameter models, while DGX Station supports up to 1-trillion-parameter models. This shift means developers can now run complex AI tasks locally, reducing latency and improving development agility.
• NVIDIA Grace Blackwell Architecture: Both systems are powered by the NVIDIA Grace Blackwell architecture, which features large unified memory (775GB on DGX Station) and petaflop-level performance. This architecture is crucial for handling the computational demands of large AI models.

Performance Enhancements

• Model Optimization: Continuous advancements in model optimization and collaboration with the open-source community have made it possible to run highly optimized models that previously required data center resources.
• NVFP4 Data Format: The NVIDIA Blackwell architecture includes the NVFP4 data format, which compresses AI models by up to 70% without sacrificing performance. This compression is particularly beneficial for reducing memory usage and improving model loading times.

System Capabilities

DGX Spark

• 100-Billion-Parameter Models: Suitable for developers looking to run large-scale AI models locally.
• Preconfigured Software: Comes with NVIDIA AI software and CUDA-X libraries, providing a plug-and-play environment for building, fine-tuning, and running AI models.
• Performance Boosts: Collaborations with the open-source community, such as llama.cpp, have resulted in a 35% performance uplift on average when running state-of-the-art AI models.

DGX Station

• 1-Trillion-Parameter Models: Ideal for enterprises and research labs that need to handle the most advanced, large-scale frontier AI models.
• GB300 Grace Blackwell Ultra Superchip: Equipped with 775GB of coherent memory and FP4 precision, enabling it to run models up to 1 trillion parameters.
• Advanced Models Supported: Includes a variety of cutting-edge models like Kimi-K2 Thinking, DeepSeek-V3.2, Mistral Large 3, Meta Llama 4 Maverick, Qwen3, and OpenAI gpt-oss-120b.

Use Cases and Impact

• Local Development and Testing: Developers can now test and develop AI models locally, reducing the need for cloud resources and improving development speed.
• Scalability to the Cloud: While these systems are designed for local use, they also provide a seamless path to scaling up to cloud infrastructure when needed.
• Accessibility for Research Labs: Frontier AI labs can leverage DGX Station's powerful capabilities without the logistical challenges of deploying rack-scale systems.

Example: vLLM Collaboration

Kaichao You, core maintainer of vLLM, noted that "NVIDIA GB300 is typically deployed as a rack-scale system. This makes it difficult for projects like vLLM to test and develop locally." The introduction of DGX Station addresses this challenge by providing a desktop solution with the compute power of a data center.

Conclusion

The NVIDIA DGX Spark and DGX Station represent a significant advancement in making high-performance AI development more accessible. By bringing petaflop-level performance to the desktop, these systems empower developers to run large-scale models locally, accelerating innovation across various industries.