NVIDIA introduced RTX Spark, a hardware and software platform designed to let developers run coding agents, AI frameworks and deployment workflows locally on one PC. The product aims to move workflows away from a split setup of laptop, cloud model API and remote VM so that coding, testing and deployment can all happen on the same device.
Hardware fundamentals and memory
RTX Spark systems are built around the NVIDIA GB10 system‑on‑a‑chip and provide large unified memory: up to 128 GB of unified memory is available. NVIDIA specifies up to 1 petaflop of FP4 AI compute, enabling local execution of coding models (for example, Qwen 3.8 Flash Next) and reducing consumption of cloud tokens.
One machine for the full development cycle
RTX Spark laptops and compact desktops are intended to let developers code, test and deploy without switching machines. The systems support Windows, including the new Windows Developer Config developed with Microsoft, which installs common development frameworks, silences notifications, improves the terminal and can optionally move the taskbar to the side. Developers can also use Windows Subsystem for Linux (WSL) to access the same dependencies and environment they will use in cloud deployments.
This setup allows work to remain in the target OS without migration. Because RTX Spark uses an NVIDIA GPU, developers can test and deploy to the largest local AI installed base: about 150 million RTX PCs, as well as the many clouds running on NVIDIA GPUs.
Performance for AI development
Development workloads have shifted from being largely CPU‑bound to GPU‑bound; today’s AI developers need GPU compute for inference (to run coding agents), for fine‑tuning and for compilation. RTX Spark pairs the NVIDIA Grace CPU with a Blackwell GPU to provide compute for inference, fine‑tuning and compilation.
NVIDIA reports that, normalized to an Intel Core Ultra X9 388H, RTX Spark delivers 4.6× performance on coding agent workloads, 2.7× on fine‑tuning and 1.8× on code compilation. The reported tests include AA‑AgentPerf‑local with Qwen3.8‑27B, Unsloth LoRA with Llama‑3.1‑8B, and Linux kernel compilation using the Phoronix Test Suite.
Familiar software stack extended to Windows on Arm
NVIDIA and Microsoft have collaborated to advance the Windows on Arm developer experience. Alongside RTX Spark, the ecosystem is shipping the first PyTorch‑CUDA build for the platform, a native‑by‑default Python 3.15 release, more than 2,000 ready‑to‑use Conda packages, and Arm‑compatible builds for over 97% of the top 15,000 Python packages.
A major advantage of developing on NVIDIA hardware is early tool and framework support: many new tools are available on NVIDIA from day one, with other hardware platforms following later. Developers also get access to CUDA‑X libraries and tools such as FlashInfer for LLM acceleration and NVIDIA NeMo libraries for model development. NVIDIA Skills provides a repository of agent instructions for using CUDA‑X libraries so teams can build GPU‑accelerated apps without deep CUDA expertise.
Local coding agents on RTX Spark
RTX Spark is optimized for local coding agents. NVIDIA worked with top agents — including GitHub Copilot, Pi, Hermes and OpenClaw — to improve integration and usability. Key enhancements include:
- seamless deployment of local models via optimized recipes
- increased security using NVIDIA OpenShell and/or Microsoft Execution Containers for OS‑level sandboxing
- hybrid prompt routers that send prompts to local models when adequate and escalate to cloud models when necessary, reducing token usage
Microsoft GitHub Copilot now supports seamless local AI via Windows ML with llama.cpp as a backend; the HydraFusion prompt router was updated to support local models; and the MAI 1.1 Coding Flash model can run locally on RTX Spark.
Pi, an open‑source coding agent with a lightweight terminal interface and support for local and cloud models, is optimized for RTX Spark and offers extensions, skills and prompt templates to fit into developer workflows. JetBrains added native support for Junie Local on RTX Spark: Junie reads code, edits files and runs tests while keeping model and code on the machine. Junie Local sizes its context window to the available GPU memory, and RTX Spark’s unified memory lets it keep more code, tool output and conversation history in context.
New Nemotron models for local agents
NVIDIA introduced Nemotron 3.5 Bolt 74B‑A9B 24GB, a large language model designed to run local agents on RTX Spark. The model uses a 74‑billion‑parameter architecture with a 9‑billion‑active‑parameter mixture‑of‑experts (MoE) design and is packed into 24 GB using 2‑bit quantization and quantization‑aware training to preserve quality. The model is intended to balance speed and capability: faster than many popular dense models while offering higher intelligence than many MoE models with a similar memory footprint.
At launch, Nemotron 3.5 Bolt 74B will be runnable via llama.cpp, vLLM, Ollama, Unsloth and LM Studio, and will be accessible through agents such as Hermes and OpenClaw. Like other Nemotron releases, Bolt 74B is fully open source — training data, weights and recipes are published so developers can fine‑tune and customize models.
Decision models, media SDKs and porting guidance
Decision models — which let agents make near‑instant choices without long reasoning loops, and can be up to 200× faster than traditional approaches for some selection tasks — are now supported in llama.cpp and Unsloth on RTX Spark and GeForce RTX GPUs.
The NVIDIA AI for Media SDKs are available for RTX Spark, including a Python wheel on PyPI supporting video super resolution, frame generation and TrueHDR; C++ SDKs are also available on NGC for RTX Spark and Windows on Arm.
To port an existing Windows application to RTX Spark, developers should assess Windows on Arm dependencies, select a native Arm64 build or use Arm64EC to migrate incrementally, and test installation, functionality, performance and NVIDIA/CUDA usage. NVIDIA’s app‑porting resources provide guidance and tools for migration planning.
Availability and next steps
RTX Spark laptops are available now. Small form factor desktops optimized for always‑on, 24×7 agent workloads are scheduled to ship in November. Developers can explore RTX Spark systems and purchasing options, and use the AI Development playbook to configure Windows, set up agents and connect NVIDIA Skills. NVIDIA also provides resources on developing local AI apps for its GPUs.
The RTX Spark offering is positioned to let developers run the full AI development lifecycle locally on a single, GPU‑accelerated machine and to tap into NVIDIA’s broad software and model ecosystem.
