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

NVIDIA’s developer blog says RTX Spark combines the GB10 system-on-a-chip with up to 128GB of unified memory so developers can run local coding agents, test, and deploy on one PC. Versus Intel Panther Lake 12Xe, NVIDIA claims 4.6x faster coding-agent performance, 2.7x faster fine-tuning, and 1.8x faster code compilation. The company partnered with Microsoft on Windows on Arm tooling, citing a first PyTorch-CUDA build, Python 3.15 native by default, over 2,000 Conda packages, and Arm-compatible versions for more than 97% of the top 15,000 Python packages.

Nemotron 3.5 Bolt 74B-A9B 24GB is positioned for local agents on Spark via llama.cpp, vLLM, Ollama, Unsloth, and LM Studio. Decision models in llama.cpp and Unsloth are said to choose actions up to 200x faster than traditional reasoning loops. Laptops are available now; small-form-factor desktops are due in November.

The split laptop/API/VM workflow is real friction. If Spark actually keeps agents, fine-tunes, and CUDA deploys on one box, that’s useful AI for developers who hate token meters for every autocomplete.

We’ll discount vendor multiples until outside benches confirm them. The software story — Windows on Arm + CUDA day-one tools — is the harder part, and Ars Technica’s Surface Ultra review already shows gaps. Still, Bolt-in-24GB and decision-model support are concrete product bets.

I’m watching independent AgentPerf numbers and whether November desktops land as always-on local agent appliances.

NVIDIA cites collaboration with GitHub Copilot, Pi, Hermes, OpenClaw, and JetBrains Junie Local, plus hybrid routers that escalate to cloud only when needed.