NVIDIA has introduced the DGX Spark and DGX Station, powerful deskside AI supercomputers designed to help developers leverage advanced open-source AI models directly from their desktops. At CES, NVIDIA showcased how these systems enable users to run models with up to 1 trillion parameters, significantly enhancing local AI capabilities. The DGX Spark, for instance, can handle 100-billion-parameter models, while the DGX Station is equipped with the Grace Blackwell architecture, offering petaflop-level performance and large unified memory. This setup allows for local development and easy scaling to the cloud. The systems are preconfigured with NVIDIA AI software and CUDA-X libraries, streamlining the development process for researchers and data scientists. The NVFP4 data format enhances model performance by compressing AI models by up to 70% without sacrificing quality. Collaborations with the open-source community, such as the integration with llama.cpp, have resulted in significant performance boosts, averaging a 35% uplift for state-of-the-art AI models. Notable AI applications include various advanced models like Qwen3 and Mistral Large 3, which can now be run efficiently from a desktop environment. NVIDIA's DGX Spark and Station support a wide range of industry-specific AI applications, from healthcare to creative workflows. They enable creators to offload demanding tasks, like video generation, achieving up to 8x acceleration compared to high-end laptops. The systems also cater to AI coding assistants, enhancing developer productivity while keeping source code secure. Industry leaders, including IBM and Hugging Face, have endorsed the shift towards local AI processing, highlighting benefits like faster iterations, data control, and interactive experiences. This trend marks a significant move towards decentralized AI capabilities, allowing developers to build and deploy models without relying on centralized infrastructure.