NVIDIA says RTX Spark can localize AI coding workflows on one PC

NVIDIA on Friday said its new RTX Spark systems are designed to let developers code, test and deploy AI software on a single machine, rather than splitting work between a laptop, cloud model APIs and a cloud VM. The company said the platform uses the NVIDIA GB10 system-on-a-chip and large unified memory to support local coding agents, AI frameworks and GPU-accelerated development workflows.

Evidence and context

In NVIDIA’s account, RTX Spark laptops and compact desktops are intended to keep the full development cycle on-device. The company said the systems offer up to 128GB of unified memory and 1 petaflop of FP4 AI compute, allowing local execution of coding models including Qwen 3.8 Flash Next and reducing cloud token consumption for coding agents.

NVIDIA framed the product as part of a shift from CPU-bound development to workloads dominated by inference, fine-tuning and compilation. It said RTX Spark combines NVIDIA Grace CPU and Blackwell GPU hardware, and compared with Intel’s Panther Lake 12Xe it delivers 4.6x performance on coding agents, 2.7x performance on fine-tuning and 1.8x performance on code compilation. The company said those figures were measured in its own developer-workload comparison, normalized to Intel Core Ultra X9 388H, using AA-AgentPerf-local with Qwen3.8-27B, Unsloth LoRA with Llama-3.1-8B and Linux kernel compilation via Phoronix Test Suite.

The company also said developers can work in Windows, including a new Windows Developer Config developed with Microsoft, or in Windows Subsystem for Linux, where they can use the same environment and dependencies expected for cloud deployment. NVIDIA said that approach is meant to reduce code migrations and keep development aligned with the target operating system.

Local agents, frameworks and Windows-on-Arm tools

NVIDIA said it is expanding the Windows on Arm developer stack with Microsoft and the broader community. The company cited the first PyTorch-CUDA build, the first Python release native by default with Python 3.15, more than 2,000 ready-to-use Conda packages and Arm-compatible versions for more than 97% of the top 15,000 Python packages.

The company also highlighted CUDA-X libraries such as FlashInfer for LLM acceleration and NVIDIA NeMo tools for AI model development. It said NVIDIA Skills are available to help agents use CUDA-X libraries without requiring deep familiarity with CUDA.

For local coding agents, NVIDIA said it has worked with GitHub Copilot, Pi, Hermes and OpenClaw. The improvements it listed include local model deployment recipes, OS-level sandboxing through NVIDIA OpenShell and/or Microsoft Execution Containers, and hybrid prompt routers that can escalate prompts from local models to cloud models when needed. NVIDIA said GitHub Copilot now supports local AI through Windows ML with llama.cpp as a backend, while HydraFusion has been updated to support local models and Microsoft’s MAI 1.1 Coding Flash can run locally on RTX Spark. JetBrains’ Junie Local is also said to have native support, with the model and code kept on the machine.

Models and media tools

NVIDIA said a new model, Nemotron 3.5 Bolt 74B-A9B 24GB, is built for local agents on RTX Spark. It described the model as a 74B-parameter, 9B-active-parameter mixture-of-experts model packed into 24GB using 2-bit quantization and quantization-aware training. The company said it will be available through llama.cpp, vLLM, Ollama, Unsloth and LM Studio, and directly through agents such as Hermes and OpenClaw. NVIDIA also said the model is open source, including training data, weights and recipes.

The company said decision models are now supported in llama.cpp and Unsloth on RTX Spark and GeForce RTX GPUs, describing them as a faster way for agents to choose actions without lengthy reasoning loops. NVIDIA also said its AI for Media SDKs are available for RTX Spark, including a Python wheel for video super resolution, frame generation and TrueHDR, with C++ SDKs available on NGC for RTX Spark and Windows on Arm.

Availability and deployment path

NVIDIA said RTX Spark is available now in laptops and will arrive in November in small form factor desktops designed for always-on agent workloads. For existing Windows applications, the company said developers should assess Windows on Arm dependencies, choose a native Arm64 build or Arm64EC for incremental migration, and then test installation, functionality, performance and NVIDIA and CUDA code paths. NVIDIA’s app-porting materials, it said, outline the migration steps and preparation for Windows on Arm development systems.