NVIDIA Introduces RTX Spark for Local AI Coding, Testing, and Deployment

NVIDIA Introduces RTX Spark for Local AI Coding, Testing, and Deployment NVIDIA has announced NVIDIA RTX Spark, a platform powered by the NVIDIA GB10 system-on-a-chip with Grace CPU and Blackwell GPU architectures, designed to let developers run AI coding agents, fine-tune models, and compile software entirely on a single machine. RTX Spark features up to 128GB of unified memory and 1 petaflop of FP4 AI compute, outperforming Intel's Panther Lake 12Xe across coding agent performance, fine-tuning, and code compilation. NVIDIA partnered with Microsoft to advance Windows on Arm, introducing native PyTorch-CUDA, Python 3.15 default support, and expanded Conda packages. Collaborations with GitHub Copilot, Pi, Hermes, OpenClaw, and JetBrains bring optimized local agent execution, hybrid prompt routing, and OS-level sandboxing. NVIDIA introduced Nemotron 3.5 Bolt 74B-A9B 24GB, an open-source MoE model optimized via 2-bit quantization for local agent execution. Hardware Capabilities and Performance RTX Spark laptops and compact desktops are built with the GB10 system-on-a-chip, pairing an NVIDIA Grace CPU with a Blackwell GPU. Offering up to 128GB of unified memory and 1 petaflop of FP4 AI compute, the platform transitions AI development from a CPU-bound workflow to a GPU-bound one. Compared to Intel's Panther Lake 12Xe, NVIDIA reports 4.6x faster performance on coding agents, 2.7x on fine-tuning, and 1.8x on code compilation. Windows on Arm and Ecosystem Integration Through a collaboration with Microsoft, the Windows on Arm developer landscape is expanded with a dedicated Windows Developer Config, the first native PyTorch-CUDA build, native Python 3.15, and over 2,000 Conda packages. Developers can choose Windows Subsystem for Linux or Windows directly, while tapping into CUDA-X libraries, FlashInfer, NeMo, and NVIDIA Skills to guide agents in GPU acceleration. Local Coding Agents and New Nemotron Model RTX Spark supports local coding agent integration with tools such as GitHub Copilot, Pi, and JetBrains Junie Local. Features include hybrid prompt routing between local and cloud models and security sandboxing via NVIDIA OpenShell and Microsoft Execution Containers. Alongside framework support for decision models in llama.cpp and Unsloth, NVIDIA introduced the open-source Nemotron 3.5 Bolt 74B-A9B, an MoE model that uses 2-bit quantization to fit into 24GB of memory for local agent workloads. Porting and Availability Developers can migrate existing software to RTX Spark by targeting Windows on Arm natively or using Arm64EC. RTX Spark is currently available in laptops, with small form factor desktop systems planned for release in November.

NVIDIA 推出 RTX Spark，实现单机完成 AI 编程、测试与部署 NVIDIA 推出面向 AI 开发的 RTX Spark 笔记本与紧凑型台式机，搭载 GB10 芯片，集成了 Grace CPU 和 Blackwell GPU。该平台提供最高 128GB 统一内存和 1 petaflop FP4 算力，让开发者无需依赖多套云端和本地环境，即可在单台设备上完成编码、测试与部署全流程。 搭载 GB10 SoC（Grace CPU 与 Blackwell GPU），统一内存最高达 128GB，算力达 1 petaflop FP4。 性能相比 Intel Panther Lake 12Xe，在编程智能体、微调和编译上分别提升 4.6 倍、2.7 倍和 1.8 倍。 联手微软优化 Windows on Arm 开发生态，带来原生 Python 3.15 和首个 PyTorch-CUDA 构建。 推出专为该平台优化的开源模型 Nemotron 3.5 Bolt 74B-A9B 24GB，并支持多款本地编程智能体。 软硬件协同与性能提升 RTX Spark 配备 Grace CPU 和 Blackwell GPU，支持高达 128GB 统一内存，能够直接在本地运行 Qwen 3.8 Flash Next 等模型。根据基准测试，其相比 Intel Core Ultra X9 388H（Panther Lake 12Xe），在编程智能体运行上带来 4.6 倍性能，微调性能达到 2.7 倍，代码编译速度提升 1.8 倍。开发者可在 Windows 或 WSL 环境中作业，并无缝适配覆盖 1.5 亿台 RTX PC 及各大云端 GPU 的部署环境。 Windows on Arm 软件生态拓展 NVIDIA 与微软合作推进 Windows on Arm 开发环境，推出首个 PyTorch-CUDA 构建、默认原生的 Python 3.15，并提供逾 2000 个即用型 Conda 包，排名前 1.5 万的 Python 包中有超过 97% 完成 Arm 适配。此外，平台提供 CUDA-X 库（如 FlashInfer、NeMo）以及 NVIDIA Skills 库，使编程智能体能直接调用 GPU 加速能力。 本地编程智能体与 Nemotron 新模型 RTX Spark 针对 GitHub Copilot、Pi、Hermes 和 JetBrains Junie Local 等智能体进行了优化，结合 NVIDIA OpenShell 或微软执行容器实现系统级沙箱隔离，并引入混合提示路由以节省云端 Token。配合硬件，NVIDIA 发布了 Nemotron 3.5 Bolt 74B-A9B 24GB 开源模型，采用 2-bit 量化 MoE 架构，兼顾运行速度与推理能力，可直接通过 llama.cpp、vLLM 及 Ollama 等工具部署。

AI agents are changing how software gets built. Developers have gone from writing every line of code by hand to directing agents that write most of the code. As a result, the current workflow is split across a laptop, a cloud model API, and a cloud VM. NVIDIA RTX Spark is built for AI development. Thanks to the NVIDIA GB10 system-on-a-chip and large unified memory, developers can run local coding agents, access AI frameworks, and code, test, and deploy on a single PC. One machine to code, test, and deploy RTX Spark laptops and compact desktops enable developers to complete all of their workflow on a single machine. Developers can choose to work in Windows, including the new Windows Developer Config, developed with Microsoft, which installs the top development frameworks, mutes notifications, enhances the terminal, and can move their task bar to the side. Or they can work in Windows Subsystem for Linux, where they get access to the same environment and dependencies they will use when deploying to the cloud. This enables developers to complete all their work in their target OS without code migrations. And since RTX Spark has an NVIDIA GPU, developers can test and deploy to the largest local AI installed base: 150 million RTX PCs, as well as virtually all the clouds in the world that run on NVIDIA GPUs. With up to 128GB of unified memory and 1 petaflop of FP4 AI compute, developers can run coding models locally, including Qwen 3.8 Flash Next, reducing cloud coding agent token consumption. The performance you need for AI development Development has gone from being a CPU-bound workload, where performance was determined mostly by the CPU, to a GPU-bound workload. AI developers today need to run inference to run their coding agents and test their apps; for fine-tuning and training models; and yes, still a bit of compilation. Thanks to the NVIDIA Grace CPU and Blackwell GPU in RTX Spark, these laptops and desktops come equipped with all the compute required for inference, fine-tuning and compilation. Compared with Intel’s Panther Lake 12Xe, RTX Spark delivers 4.6x performance on coding agents, 2.7x performance on fine-tuning, and 1.8x performance on code compilation. Figure 1. NVIDIA developer-workload comparison, with Intel Core Ultra X9 388H normalized to 1.0. Tests: AA-AgentPerf-local with Qwen3.8-27B; Unsloth LoRA with Llama-3.1-8B; and Linux kernel compilation using Phoronix Test Suite A familiar software stack, expanded for Windows NVIDIA partnered with Microsoft to advance the Windows on Arm developer landscape. Together with RTX Spark, the community is releasing multiple new tools, including the first PyTorch-CUDA build, the first Python release that is native by default with Python 3.15, over 2,000 ready-to-use Conda packages, and Arm-compatible versions for more than 97% of the top 15,000 Python packages. A key advantage of developing on NVIDIA hardware is access to the largest ecosystem of developer tools, from coding agents, to AI models, to frameworks. Because NVIDIA GPUs are the most widely used platform for AI development, new tools and frameworks typically work on NVIDIA from day one, while support for other hardware follows weeks later. There’s also access to NVIDIA CUDA-X libraries, including tools such as FlashInfer for LLM acceleration and NVIDIA NeMo libraries tools for AI model development. NVIDIA also provides a repository of NVIDIA Skills that teach agents how to use CUDA-X libraries, enabling developers to build GPU-accelerated applications without requiring deep familiarity with CUDA. Figure 2. RTX Spark’s ecosystem of agents, models, frameworks, and CUDA-X libraries Local coding agents come to RTX Spark RTX Spark devices are ideal for local coding agents. NVIDIA has collaborated closely with the top agents, including GitHub Copilot, Pi, Hermes, and OpenClaw, to improve them and make them easier to use. The improvements include: Seamless deployment of local models with optimized recipes Enhanced security with NVIDIA OpenShell and/or Microsoft Execution Containers, delivering OS-level sandboxing Hybrid prompt routers between local and cloud models, so that prompts get sent to local models if they are good enough, or escalated to cloud if needed, reducing token consumption Microsoft GitHub CoPilot now enables seamless local AI through Windows ML, which is now supported with llama.cpp as a backend. The prompt router HydraFusion has been upgraded to also support local models. And the new 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 also optimized for RTX Spark. Its extensions, skills, and prompt templates give developers control over how the agent fits into their workflow, from model selection to the tools it can use. JetBrains has launched native support for Junie Local on RTX Spark . The coding agent reads code, edits files, and runs tests, with the model and code staying on the machine. Junie Local sizes its context window to the memory available to the GPU, so RTX Spark unified memory lets it keep more code, tool output, and conversation history in context. New Nemotron models offer more options for local agents Nemotron 3.5 Bolt 74B-A9B 24GB is a new LLM purpose-built for running local agents on RTX Spark. This model is a 74B parameter, 9B-active-parameter mixture-of-experts (MoE) model, packed into 24GB using 2-bit quantization with advanced quantization-aware training techniques to retain quality. It aims to achieve the optimal balance between speed and intelligence, being faster than most popular dense models while smarter than most popular MoE models with a similar memory footprint. Users will be able to run Nemotron 3.5 Bolt 74B at launch through llama.cpp, vLLM, Ollama, Unsloth and LM Studio. It will also be directly accessible via agents such as Hermes and OpenClaw. Like other Nemotron models, Bolt 74B is fully open source, including training data, weights and recipes, so developers can fine-tune and customize their own agents. Decision models now supported in top frameworks Decision models are the latest breakthrough in AI development. They give AI agents a near-instant way to make decisions, without lengthy reasoning loops. They can choose what to do next, from classifying requests, to selecting the next action, up to 200x faster than traditional agents. Decision models are now supported in llama.cpp and Unsloth on RTX Spark and GeForce RTX GPUs. The NVIDIA AI for Media SDKs are now available for RTX Spark, including Python wheel with support for video super resolution, frame generation, and TrueHDR. The wheel is available on PyPI, with C++ SDKs also available on NGC for RTX Spark and Windows on Arm. Bring your Windows applications to RTX Spark To bring an existing application to RTX Spark, first assess dependencies for Windows on Arm. Choose a native Arm64 build or use Arm64EC to migrate components incrementally. Test installation, functionality, performance, and the NVIDIA and CUDA paths your app uses. NVIDIA’s app-porting resources explain the migration options and tools, including how to start preparation on an existing Windows on Arm development system. Find your RTX Spark and start building RTX Spark is available now in laptops, and coming in November to small form factor desktops that are ideal for always on, 24×7 running agents. Explore RTX Spark systems to find your favorite and see available purchasing options. Once you get your own, get started with the AI Development playbook to configure Windows, set up your agent, and connect NVIDIA Skills. And learn about how you can get started developing Local AI apps for NVIDIA GPUs.