NVIDIA's RTX Spark puts the full coding-agent workflow on one machine

NVIDIA's RTX Spark is available now in laptops, with compact desktops due in November. It pairs the GB10 system-on-a-chip (Grace CPU plus Blackwell GPU) with up to 128GB of unified memory and 1 petaflop of FP4 AI compute, and the pitch is a single machine for the whole agent-driven development loop: run coding agents locally, fine-tune, compile, test and deploy without a cloud VM in the middle.

One machine, vendor numbers

The workflow argument is this: agent-era development is GPU-bound, not CPU-bound. You're running inference for the coding agent, doing some fine-tuning, and still compiling code. NVIDIA measured its platform against Intel's Core Ultra X9 388H (Panther Lake 12Xe) and, according to the company, delivers 4.6x the performance on coding agents (AA-AgentPerf-local with Qwen3.8-27B), 2.7x on fine-tuning (Unsloth LoRA with Llama-3.1-8B) and 1.8x on Linux kernel compilation. Those are NVIDIA's numbers, from NVIDIA's own test setup, and independent results are not out yet.

The unified memory is the practical part: with 128GB, models like Qwen 3.8 Flash Next run locally, which reduces the token bill for cloud coding agents. And because the target is an NVIDIA GPU, what you test on the laptop is what deploys to the 150 million RTX PCs and to the clouds that run NVIDIA silicon. No code migration between dev machine and target.

A Windows on Arm stack that finally ships CUDA

NVIDIA says the usual blocker for Windows on Arm development, the missing CUDA path, is the thing it fixed with Microsoft: the first PyTorch-CUDA build for the platform, Python 3.15 native by default, more than 2,000 ready-to-use Conda packages, and Arm-compatible versions for over 97% of the top 15,000 Python packages. There's also a joint Windows Developer Config that pre-installs the main frameworks and quiets notifications.

If Windows on Arm is not your environment, WSL on the same machine gives you the same dependencies you'd deploy with. For existing applications, the migration path is native Arm64 builds or Arm64EC to move components incrementally, with NVIDIA's app-porting docs covering the tools.

Local agents, and a new 74B model for them

The ecosystem side is where most of the news sits. NVIDIA worked with GitHub Copilot, Pi, Hermes and OpenClaw on local-model support. Copilot runs local AI through Windows ML with llama.cpp as backend, and the new MAI 1.1 Coding Flash model runs on the device. JetBrains launched native Junie Local support, with the context window sized to whatever the GPU's memory allows.

The new model is Nemotron 3.5 Bolt 74B-A9B: a 74B-parameter mixture-of-experts with 9B active parameters, packed into 24GB with 2-bit quantization. It ships at launch through llama.cpp, vLLM, Ollama, Unsloth and LM Studio, and NVIDIA says it is fully open source, including training data and recipes. Decision models, the routing-style models agents use to pick the next action, are now supported in llama.cpp and Unsloth, with NVIDIA claiming up to 200x faster action selection than traditional reasoning loops.

The practical read

If your workflow is already split across a laptop, a cloud model API and a cloud VM, RTX Spark is NVIDIA's answer to collapsing that into one box. The buy depends on two things you should check first: whether your toolchain is actually Arm64-ready (the 97% coverage number leaves 3%), and whether the local-model savings beat your current cloud spend for your workload. Some may call the 200x decision-model number hype until independent benchmarks show up, and the 4.6x headline is vendor-measured. The hardware, the memory ceiling and the CUDA-on-Windows-on-Arm stack are the durable part of this announcement.