NVIDIA announced the availability of NVIDIA RTX Spark laptops on 16 October 2026, featuring a Blackwell and Grace superchip with up to 128GB of unified memory and 1 petaflop of FP4 AI performance. (FP4 AI performance: a specific precision format used for calculations to power artificial intelligence tasks.) (unified memory: a shared memory architecture where both the central processing unit and graphics processing unit can access the same pool.) The platform launched across ASUS, Dell, HP, Lenovo, Microsoft, and MSI, targeting creative professionals and AI developers who need to run frontier models locally. If you currently rely on cloud-based AI for inference or use standard Windows laptops with limited shared memory, these new local execution capabilities change your hardware requirements.
Is this for you?
If you are a creative professional or AI developer, the 128GB unified memory allows you to run large models locally instead of using the cloud.
- If you use Adobe Premiere and Photoshop, the RTX Spark platform provides native TensorRT acceleration for 8K video editing.
- If you develop for Windows on Arm, the platform offers a native Python build and the first PyTorch-CUDA build.
- If you work with large 3D scenes, the 90GB rendering limit determines if you can work locally or need a cloud service.
- Skip this if you are a budget-conscious student as the starting price is $2,599.
- Skip this if you only use cloud-based AI as the hardware is designed for local execution.
What you can do now
- If you have 3D scene files, verify if they are under the 90GB limit for local rendering.
- If you are a Windows on Arm developer, configure a native Python environment on the RTX Spark hardware.
- If you use the standard Conda installer, compare its installation speed against the native Conda installer.
Key figures
- Availability date
- 16 October 2026
- AI compute performance
- 1 petaflop of FP4 AI performance
- Unified memory capacity
- Up to 128GB
- Starting price
- $2,599 at launch
- Primary hardware
- Blackwell RTX GPU and NVIDIA Grace CPU superchip
NVIDIA Blackwell and Grace superchip hardware foundation
NVIDIA RTX Spark laptops utilize a single superchip combining an NVIDIA Blackwell RTX GPU and an NVIDIA Grace CPU to deliver high-performance computing in a portable form factor. The architecture provides up to 128GB of unified memory—a shared memory architecture where both the CPU and GPU can access the same pool of memory, enabling larger models to run locally—alongside up to 1 petaflop of FP4 AI performance. MediaTek collaborated with NVIDIA on the custom CPU design to achieve best-in-class power efficiency and connectivity in these slim designs.
The superchip architecture allows for the simultaneous execution of demanding creative
The superchip architecture allows for the simultaneous execution of demanding creative projects and large-scale local AI models without memory bottlenecks. By consolidating the GPU and CPU onto a single chip with shared memory, NVIDIA enables the Blackwell/Grace superchip to operate within a portable form factor suitable for a messenger bag. This hardware consolidation moves the boundary of what portable Windows PCs can compute compared to traditional discrete GPU setups.
Local AI agents and 1 petaflop of FP4 performance
Local AI agents like Hermes, OpenClaw, Perplexity, Pi, and WorkBuddy are supported on the RTX Spark platform to turn the PC into an everyday teammate. Users can deploy these agents using guided setup and easy installers that do not require a command line. Dillon Rolnick, CEO at Nous Research, stated they brought Hermes to RTX Spark so people can put open AI models and agents to work locally on a Windows laptop with greater control over their data.
NVIDIA models and tools
NVIDIA models and tools, including the 74-billion-parameter Nemotron 3.5 Bolt 74-B-A9B 24GB model, provide more capable local AI for supported RTX Spark configurations. The 1 petaflop of FP4 AI performance—a specific precision format used for AI calculations—enables these agents to perform complex tool calling and multistep tasks. Because the hardware supports local execution, users can avoid the latency and costs associated with cloud-based AI inference.
Adobe Creative Cloud and 8K video editing
Adobe Premiere and Photoshop offer native RTX Spark support through beta releases available at launch, featuring NVIDIA TensorRT acceleration. (TensorRT acceleration: a software suite used to optimize and speed up deep learning inference on graphics hardware.) Creators can perform 8K 4:2:2 video editing and render large 3D scenes up to 90GB, even when the laptop is unplugged. Adobe’s Deepa Subramaniam stated that combining NVIDIA’s technology with Adobe’s applications delivers enhanced experiences for creatives who need to bring their ideas to life with greater power and speed.
NVIDIA Broadcast 2.3 provides additional tools for creators to remove background noise
NVIDIA Broadcast 2.3 provides additional tools for creators to remove background noise from existing recordings and shape voices using Studio Voice Mic Profiles. ComfyUI users can also utilize new RTX Video nodes to turn 15 frames-per-second (fps) AI-generated video into smoother 60 fps footage and convert it to high dynamic range without requiring per-use cloud inference charges. The 90GB rendering limit serves as a hard boundary for local 3D scenes; users requiring more would still need to rely on a cloud rendering service.
Windows on Arm and Python developer tools
Windows on Arm developers gain the first PyTorch-CUDA build for the RTX Spark platform alongside a Python build that is native by default. (PyTorch-CUDA build: a version of the machine learning library configured to run on specific graphics processing unit hardware.) The platform also includes a native Conda installer and supports 97% of the top 15,000 Python packages. (Conda: a package and environment management system for programming languages like Python.) These tools provide a familiar hybrid environment that bridges local development with cloud-based deployment via the Windows Subsystem for Linux.
High-end gaming with DLSS 5
High-end gaming is supported through NVIDIA Blackwell RTX graphics and DLSS 5 technology, which includes 3D-Guided Neural Rendering to enhance lighting and material details. NBA 2K27 demonstrates the platform’s capability, achieving over 100 frames per second at 1440p resolution. A new 2x faster AI model, available 16 October 2026, boosts DLSS 5 frame rates by up to 1.4x.
RTX Spark laptop lineup and pricing
The available laptop lineup includes the following models:
- ASUS
- ProArt P14 and ProArt P16
- Dell
- XPS 16
- HP
- OmniBook Ultra and OmniBook X 14
- Lenovo
- Yoga Pro 9n
- Microsoft
- Surface Laptop Ultra
- MSI
- Prestige N16 Flip AI+ and Prestige N16
Pricing for these laptops starts at $2,599, with specific configurations and costs varying by region.
Who this is for: Adobe Premiere and Photoshop users
Adobe Premiere and Photoshop users who need to edit 8K 4:2:2 video and render large 3D scenes locally should use the RTX Spark platform to access native TensorRT acceleration. AI developers working on local LLMs can utilize the 128GB unified memory to identify which frontier AI models can fit into the local memory for deployment. Windows on Arm developers can integrate the first PyTorch-CUDA build into their local development workflow to build for Arm-based hardware.
Who this is not for
Budget-conscious students should skip this offering as the $2,599 starting price is for premium hardware; these users should look for entry-level NVIDIA RTX laptops instead. Users only interested in cloud-based AI can ignore this news, as the hardware is designed for local execution, and they should instead consult cloud AI service documentation.
Scenario for a freelance animator
Henrik runs a freelance 3D animation studio and owns an ASUS ProArt P16 to handle his client’s work. Because the laptop features 128GB of unified memory, Henrik can keep his 3D modeling software open while simultaneously running a local Hermes agent to help generate textures and assets. This allows him to complete his work on location without needing a desktop or a cloud-based rendering farm.
If Henrik’s 3D scene size increased to 100GB, the calculation would flip because the source specifies a 90GB limit for local 3D scene rendering. In that case, Henrik would need to switch to a cloud rendering service for that specific project.
Applications for RTX Spark
| Use Case | Who it serves |
|---|---|
| Running frontier AI models locally | Data Scientists and AI Researchers |
| Broadcasting with AI-enhanced voice profiles | Streamers and Content Creators |
| Simultaneous creative work and AI agent usage | Multi-tasking Designers |
| Serving AI inference to other PCs on a home network | Home Lab Enthusiasts |
| Native Python development on Windows on Arm | Arm-based Developers |
What RTX Spark replaces
Cloud inference costs are reduced because the 128GB of unified memory allows for local execution of large language models and AI agents. Command-line only AI setup is replaced by guided setup and easy installers for agents like Hermes and OpenClaw. Multi-PC inference setups are replaced by NVIDIA PAIR, which lets the laptop serve AI inference to other compatible Windows, Linux, and macOS computers on a home network.
Unconfirmed details and limitations
NVIDIA did not state what specific hardware components are removed or changed for existing users who might upgrade. The source did not explain how to use the Windows Developer Config tool for the initial setup of the RTX Spark device. Additionally, the source did not specify any known hardware limitations or performance bottlenecks for the new superchip.
NVIDIA RTX Spark: what to check, compare and watch
- Adobe Premiere Pro
- The RTX Spark hardware enables new TensorRT acceleration for features within Adobe Premiere Pro. Adobe Premiere and Photoshop users: compare the rendering speeds of the current Premiere Pro version against the new RTX Spark optimized version.
- 3D scene files
- RTX Spark allows users to render 3D scenes up to 90GB locally. Adobe Premiere and Photoshop users: verify if current 3D scene files are under the 90GB limit for RTX Spark rendering.
- Substance 3D update
- The source mentions that Substance 3D support is coming soon for the RTX Spark platform. Adobe Premiere and Photoshop users: monitor the official Adobe blog for the Substance 3D release date for RTX Spark.
- GPU VRAM
- RTX Spark's 128GB unified memory provides a different architecture than standard GPU VRAM. AI developers working on local LLMs: compare the performance of 128GB unified memory against standard discrete GPU VRAM for model inference.
- Python environment
- RTX Spark provides a Python build that is native by default for Windows on Arm. Windows on Arm developers: configure a native Python environment on the RTX Spark hardware for Arm-based development.
- Standard Conda installer
- RTX Spark offers a native Conda installer compared to standard versions. Windows on Arm developers: compare the installation speed of the native Conda installer against the standard installer.
NVIDIA RTX Spark: questions the original report leaves open
- How do you get it or turn it on?
- The original report did not say how to turn it on.
- What does it require, and is anything removed or changed for existing users?
- The original report did not say what is removed or changed for existing users.
- Are there known problems or limitations?
- The original report did not say what the known problems or limitations are.
- How to use Windows Developer Config for RTX Spark?
- The original report did not say how to use Windows Developer Config.
Original reporting: NVIDIA RTX Spark PCs Available Today — A New Beginning for Windows PCs (NVIDIA), on 16 October 2026.