NVIDIA Corporation (NASDAQ: NVDA) has expanded its RTX Spark developer stack around a simple proposition: run coding agents, test applications and prepare deployments from the same Windows-based machine. The company said the systems combine its GB10 system-on-a-chip with as much as 128GB of unified memory for local AI development.
NVIDIA RTX Spark is aimed at a software workflow increasingly split among a developer laptop, a cloud model service and a cloud virtual machine. NVIDIA said developers can use Windows or Windows Subsystem for Linux, which lets them work with the same operating system environment and dependencies that may be used in cloud deployment. That is intended to reduce code migration between local development and a target environment.
The company said the systems provide up to one petaflop of FP4 AI compute and can run local coding models, including a model it identifies as Qwen 3.8 Flash Next. NVIDIA also cited its own comparison against an Intel Panther Lake 12Xe configuration, reporting 4.6 times the coding-agent performance, 2.7 times faster fine-tuning and 1.8 times faster code compilation. Those measurements are NVIDIA benchmarks rather than independently verified results.
For IT teams, the platform change is as much about the surrounding software as the hardware. NVIDIA said it worked with Microsoft Corporation (NASDAQ: MSFT) to extend the Windows on Arm developer stack with a PyTorch-CUDA build, a Python 3.15 release that is native by default, more than 2,000 Conda packages and Arm-compatible versions for more than 97% of the top 15,000 Python packages. The source does not provide an independent compatibility audit.
NVIDIA also described updated support for local coding agents, including tools from GitHub Copilot, Pi, Hermes and OpenClaw. It said optimized recipes can simplify local model deployment, while hybrid prompt routers can direct a request to a local model when it is suitable and escalate it to a cloud model when needed. That approach could reduce cloud token use, but the company did not quantify savings for a typical enterprise workload.
Security is a practical constraint when AI agents can use developer tools and access source code. NVIDIA said RTX Spark can use NVIDIA OpenShell or Microsoft Execution Containers for operating-system-level sandboxing. JetBrains’ Junie Local can read code, edit files and run tests while keeping the model and code on the machine, according to the company. Those controls do not eliminate the need for enterprise access policies, code review and software-supply-chain protections.
The hardware release also introduces Nemotron 3.5 Bolt 74B-A9B 24GB, a 74-billion-parameter mixture-of-experts model with nine billion active parameters that NVIDIA said is optimized for local agents. It is available through several inference and local-model tools, and NVIDIA said its weights, training data and recipes are open source for developers who want to customize agents.
RTX Spark systems are available in laptops now, with small-form-factor desktops due in November for always-on agents, NVIDIA said. For platform-engineering teams, the immediate benefit is the option to put more coding-agent inference, testing and experimentation on a controlled local machine. The trade-off remains operational: organizations will need to decide which work belongs on developer endpoints and which still requires centralized infrastructure, governance and monitoring.