NVIDIA Launches RTX Spark PCs for AI Development

NVIDIA has announced RTX Spark, a developer system pairing the GB10 system-on-a-chip—built on Grace CPUs and Blackwell GPUs—with up to 128GB of unified memory and 1 petaflop of FP4 AI compute. Available now in laptops and arriving in November in compact 24x7 desktops, the platform lets engineers build, test, and deploy software locally rather than splitting workflows across separate machines and cloud APIs. Compared to Intel's Panther Lake 12Xe (specifically the Core Ultra X9 388H), NVIDIA reports RTX Spark reaches 4.6x faster coding-agent speeds on AA-AgentPerf-local using Qwen3.8-27B, 2.7x faster fine-tuning performance running Unsloth LoRA with Llama-3.1-8B, and 1.8x faster code compilation on Linux kernel benchmarks.

To strengthen the Windows on Arm environment, NVIDIA collaborated with Microsoft to introduce the first native PyTorch-CUDA build, default native Python 3.15, more than 2,000 Conda packages, and Arm compatibility for over 97% of the top 15,000 Python packages. The ecosystem incorporates tools like FlashInfer for LLM acceleration, NeMo libraries, and sandboxing options via NVIDIA OpenShell or Microsoft Execution Containers. Developers can target 150 million RTX PCs alongside cloud deployments, utilizing agents like GitHub Copilot, Pi, Hermes, OpenClaw, and JetBrains Junie Local. The update also adds support for decision models in llama.cpp and Unsloth to accelerate agent action choices up to 200x over standard reasoning loops, plus HydraFusion prompt routing and NVIDIA AI for Media SDKs featuring video super resolution, frame generation, and TrueHDR via PyPI and NGC.

Along with supporting local models like Qwen 3.8 Flash Next and MAI 1.1 Coding Flash, NVIDIA introduced Nemotron 3.5 Bolt 74B-A9B 24GB. This open-source 74B-parameter mixture-of-experts model uses 9B active parameters and 2-bit quantization to fit into 24GB of memory while running across llama.cpp, vLLM, Ollama, Unsloth, and LM Studio. For practitioners, this setup minimizes cloud token costs and eliminates cross-platform migrations by allowing local execution of memory-heavy context windows and agentic workflows directly inside Windows or Windows Subsystem for Linux.