NVIDIA has recently launched the Underdog application, aimed at simplifying the management and operation of local AI models. The application is specifically designed for RTX Spark devices, enabling AI tasks to be processed on the computer side through a dedicated inference engine without the need for cloud connectivity, ensuring that related data remains on the user's device. The NVIDIA RTX Spark laptops were released on October 16, featuring Arm architecture-based system-on-chip technology and running Windows, primarily targeting high-load AI applications such as local large language models. Currently, products from ASUS, Dell, HP, Lenovo, Microsoft, and MSI have been launched, with Acer models also in preparation.
According to reports, Underdog employs the Husky inference engine, designed for specific models, allowing users to run local AI models directly on RTX Spark devices. Since the inference process does not rely on cloud services, user data does not need to be uploaded to external servers, reducing the privacy risks associated with data leaving local devices.
Underdog Application Enhances Local AI Model Management Efficiency
NVIDIA also announced performance updates for local inference tools. The llama.cpp now includes MTP inference decoding support for Qwen 3.8 Flash Next. When running on devices equipped with RTX Pro 6000 Blackwell, the processing speed for end-to-end agent tasks can be improved by approximately 1.4 times. Additionally, NVIDIA stated that Nemotron Bolt E4B 4B, GLM, DeepSeek, Qwen, MiniMax, and FLUX 3 Klein diffusion models will receive initial support.
This means that the relevant models can be used on compatible devices immediately after their release, without users needing to wait for subsequent compatibility updates. Underdog, along with the aforementioned support plan, further enhances the local AI software environment for RTX Spark devices.
Disclosure: This article contains links to partner products. If you purchase through these links, TechRitual may receive a commission, but this does not affect product evaluations and recommendations. For more details, please refer to our privacy policy.