At the GTC Taipei event, NVIDIA introduced NVIDIA RTX Spark, a new class of Windows personal computers designed for local artificial intelligence agents.
These local agents can interact with applications, generate content, and manage multi-step tasks directly on the device. Consequently, users can run these processes privately without relying on cloud-based infrastructure.
In addition to the hardware announcement, the company introduced several software updates to expand local agent capabilities. Specifically, these updates target both the consumer RTX platform and the enterprise DGX systems. Meanwhile, developers are adopting these tools to build more secure on-device applications.
Hardware Specifications of NVIDIA RTX Spark
The newly announced NVIDIA RTX Spark features up to 1 petaflop of AI compute power.
Furthermore, the system includes 128GB of unified memory to handle the demanding processing requirements of local agents. This hardware configuration is designed for slim Windows laptops as well as efficient desktop computers.
Alongside these consumer devices, the company introduced the DGX Station for Windows. This deskside supercomputer brings data-center-class graphics and central processing units to professional desktop environments. Notably, this system combines high-performance hardware with Windows manageability and security features.
Security Framework and Open Source Integrations
To address privacy concerns, NVIDIA is partnering with Microsoft to deliver a secure platform for local agents. This collaboration introduces the OpenShell runtime, which runs on new Windows security primitives. As a result, users can define strict policies regarding what data local agents can access.
Furthermore, open-source developer projects like Hermes Agent and OpenClaw are integrating these security features. These integrations allow the agents to execute tasks across different apps while keeping personal data disguised. Consequently, users can perform semantic file searches and generate media locally with higher security.
Software Optimizations and Performance Gains
The company also announced significant performance updates for open-source AI models. Specifically, collaboration with the llama.cpp community has enabled multi-token prediction capabilities. This technique allows a smaller draft model to propose multiple tokens that the target model verifies in a single pass.
Consequently, these optimizations deliver up to a 2x throughput increase on Qwen 3.6 and 3.5 models. Meanwhile, multi-GPU setups running llama.cpp now support tensor parallelism for improved memory and compute efficiency. In addition, ComfyUI receives updates that double performance on dual-GPU configurations.
Creative Application Updates and Future Outlook
Adobe is currently rearchitecting its Premiere and Photoshop applications specifically for NVIDIA RTX Spark. These updates will allow creative tools like Generative Fill and Generative Extend to run up to 2x faster. Moreover, the new video pipeline will utilize the system’s unified memory and Blackwell architecture.
In addition, Blender is integrating DLSS 4.5 Ray Reconstruction as a new denoiser this fall. This integration will turn the path-tracing viewport into an interactive, real-time viewer for 3D artists. Finally, the NVIDIA Broadcast 2.2 update is graduating its Studio Voice feature out of beta today.
