Code, Test, and Deploy on One Machine with NVIDIA RTX Spark | NVIDIA Technical Blog

This post introduces NVIDIA RTX Spark, a new platform designed for local AI development on a single PC. It highlights the integration of the GB10 SoC with up to 128GB of unified memory to run coding agents, test applications, and deploy models locally. Performance benchmarks show significant improvements over Intel's Panther Lake 12Xe for coding agents, fine-tuning, and compilation. The post details expanded Windows on Arm software stack support, including the first PyTorch-CUDA build and extensive Python package compatibility. It also introduces Nemotron 3.5 Bolt 74B, a quantized MoE model optimized for local agents, and the integration of decision models into frameworks like llama.cpp and Unsloth for faster agent decision-making. AI for Media SDKs are also now available on RTX Spark.