NVIDIA DGX Spark: Local AI at DevSparks Pune 2026
NVIDIA solutions architect presents on local AI systems at DevSparks Pune.
Key Highlights
NVIDIA and RP Tech demoed NVIDIA DGX Spark at DevSparks Pune 2026, showcasing local AI for enhanced data privacy and efficient AI model execution.
NVIDIA, in collaboration with RP Tech, an NVIDIA partner, highlighted the growing importance of local AI systems at the DevSparks Pune 2026 masterclass session. The session, titled ‘Introduction to NVIDIA DGX Spark and Building a VSS Agent,’ was led by Ajay Kumar Kuruba, Senior Solutions Architect at NVIDIA.
The masterclass focused on NVIDIA DGX Spark, NVIDIA’s desktop-class AI compute system built on the Blackwell architecture. Participants learned how to construct a Video Search and Summarization (VSS) agent, which converts raw video into searchable insights using vision language models, all while maintaining strong data privacy control.
Kuruba explained that data security and privacy are key reasons to use compact, local systems capable of running models at the same level as larger systems. NVIDIA DGX Spark features a GB10 GPU, a 20-core ARM processor, and 128 GB of shared memory between the CPU and GPU, connected via NVLink.
The session also covered the software stack, from CUDA drivers and the NVIDIA Container Toolkit at the kernel level to TensorRT-LLM and vertical-specific SDKs. The Container Toolkit addresses library compatibility issues by containerizing the entire stack, ensuring a consistent environment.
Kuruba emphasized that NVIDIA’s architectures and frameworks do not extract user data to train models. The session also covered quantization, with Blackwell architecture performing multiplications at the FP4 level and accumulating results at FP8.
The VSS agent was presented as an open-source blueprint that processes input from live video streams and computer vision pipelines. It uses DeepStream for chunking, sampling, and preprocessing, and passes the output to a Cosmos vision language model that generates summaries, alerts, and safety violation reports.
NVIDIA DGX Spark is designed for teams needing a dedicated local system to run models under 10 billion parameters, free from cloud dependency. The masterclass demonstrated that it is a practical option for specific requirements.