NVIDIA is making an aggressive push to bring powerful, privacy-focused AI agents directly onto personal computers, announcing a new class of RTX-powered Windows PCs, expanded developer tools and deep integrations with partners including Microsoft and Adobe.
At GTC Taipei during Computex, the company introduced NVIDIA RTX Spark systems, new security-focused agent infrastructure for Windows called OpenShell, and a series of software optimisations aimed at making locally running AI agents faster and more capable across RTX and DGX hardware.
NVIDIA positions RTX Spark as a new class of Windows PC built specifically for personal AI agents. Each system is designed to run agents locally, with up to 1 petaflop of AI compute and 128GB of unified memory to handle the intensive workloads involved in multi-step reasoning, content generation and automated workflows.
The systems target slim laptops with all-day battery life as well as ultra-efficient desktops, combining AI, creation and gaming capabilities. NVIDIA says RTX Spark is meant to turn a PC from a conventional tool into more of a “teammate” for everyday tasks.
Alongside the hardware, NVIDIA and Microsoft are collaborating to build out a secure platform for on-device agents in Windows. That stack combines new Windows security primitives with NVIDIA’s OpenShell runtime to tackle long-standing concerns about running powerful agents directly on a user’s primary PC.
According to NVIDIA, the new Windows-level primitives cover identity, containment, policy and end-to-end security for building and running native agents. OpenShell then adds:
Policy controls that let users specify what agents are permitted to do
Intelligent routing of queries to local models based on user privacy preferences
Mechanisms to disguise personal information in queries that must go to cloud models
This combined security and privacy layer is being adopted by leading agent developers such as Hermes Agent and OpenClaw in their new Windows applications. These apps aim to expose on-device agents that can operate across Windows software, manage cross-app workflows, generate images and video, build plug-ins and apps, and semantically search local files, all under user-defined constraints.
For professionals who need data-center-class performance in a desktop footprint, NVIDIA also introduced DGX Station for Windows. The company describes it as a deskside AI supercomputer that pairs a data-center-class GPU with a high-end CPU in a Windows-based system designed for manageability, security and compatibility in enterprise environments.
NVIDIA is also working on the software side to speed up the open models and agent frameworks that power local assistants.
Through collaboration with the llama.cpp community, NVIDIA has enabled features such as multi-token prediction (MTP), a speculative decoding technique that uses a smaller draft model to propose multiple tokens at once, which the target model then validates in a single pass. Combined with optimisations like programmatic dependent launch, NVIDIA reports:
2x performance on Qwen 3.6 and 3.5 27B models
1.6x performance on Qwen 3.6 and 3.5 35B models
These improvements are available through the llama.cpp web UI and LM Studio.
For Linux-focused developers and always-on local agents, NVIDIA is promoting DGX Spark as its most capable personal agent system in a Linux environment. DGX Spark combines large unified memory, high compute performance and compatibility with the CUDA ecosystem.
The latest DGX Spark OS release introduces a streamlined out-of-the-box experience with an updated NemoClaw installer and faster inference on leading agentic models. NemoClaw is now available for NVIDIA RTX and DGX PCs on Linux and via Windows Subsystem for Linux (WSL), offering automatic sandboxing and added support for Hermes Agent when deploying local agents.
NVIDIA has also worked with the vLLM community to further accelerate inference for agents. New optimisations in vLLM, along with new NVFP4 checkpoints for Qwen 3.6 35B, deliver a reported 2.6x performance increase on DGX Spark compared with previously available NVFP4 checkpoints from Unsloth. The changes include kernel-level improvements, mixed-precision execution and CUDA Graph support for MTP. A vLLM blog provides a walkthrough of serving NVFP4 mixture-of-experts models on DGX Spark, covering unified memory tuning and a Nemotron 3 Super reference setup.
NVIDIA is also expanding agent capabilities through a partnership with H Company. H Company’s “computer-use harness” allows agents to interact with a PC similarly to a human user, by seeing the screen and controlling the mouse and keyboard, including in applications that don’t expose APIs. NVIDIA says this capability is coming soon to RTX and DGX PCs with local model support.
To make that practical on consumer and developer hardware, NVIDIA has worked with H Company to quantise its Holo Computer Use models and to accelerate the harness, achieving a 2x speedup on NVIDIA GPUs while cutting memory usage by 35%. The models are already available for download, and the Holo Desktop app is described as coming soon.
On the creative side, NVIDIA is partnering with Adobe to rearchitect key Creative Cloud applications for RTX Spark. Adobe Premiere Pro and Photoshop are being tuned to take advantage of RTX Spark’s unified memory, next-generation GPU and NVIDIA TensorRT software.
In Photoshop, Firefly-powered Generative Fill is among hundreds of accelerated tools, and the next-generation engine will be optimised for GPU-accelerated compositing with live filters, HDR and modern natural brushing.
In Premiere Pro, a new video pipeline is designed to use unified memory, a Blackwell GPU and TensorRT for real-time editing and colour correction, accelerated AI effects and more efficient rendering of complex timelines.
Adobe Substance 3D Painter and Stager will run natively on RTX Spark for smoother, more responsive 3D texturing and scene creation workflows.
NVIDIA says these changes can deliver up to 2x faster AI, editing, colouring and effects across creative workflows. Adobe also plans to extend Premiere Pro and Photoshop so users can work with Windows agents inside the apps, treating AI as a collaborative teammate in creative projects. Updates for these Adobe tools are expected to roll out in step with RTX Spark availability.
Beyond core agents and creative apps, NVIDIA is updating several pieces of its broader RTX software stack:
NVIDIA Broadcast 2.2 graduates the Studio Voice feature from beta. Studio Voice applies AI processing to make virtually any microphone sound studio-grade, now running on GeForce RTX 3060 GPUs and above with better performance. Broadcast adds Elgato Stream Deck integration and configurable keyboard shortcuts.
Project G-Assist gains Stream Deck support via the Elgato MCP Server, enabling AI assistant controls in streaming setups.
Blender Cycles is integrating DLSS 4.5 Ray Reconstruction as a new denoiser. This transforms the path-tracing viewport into a near-final, interactive viewer so artists can navigate scenes with close-to-final render quality. The feature is slated to ship with Blender 5.3 this fall, alongside RTX Spark.
RTX Video Frame Generation will launch with RTX Spark as an AI effect that can double or quadruple video frame rates in real time. This is aimed at smoothing out typical 15–20 fps outputs from AI video models. It will be available as a Python wheel and a ComfyUI node, letting AI artists generate at low fps and then interpolate up to higher frame rates.
NVIDIA is also highlighting partner and ecosystem moves around RTX Spark:
ASUS ProArt creator laptops will ship with Black Forest Labs’ FLUX.2 Klein 4B image model preinstalled via the MuseTree app. The model is optimised in NVFP4 format with NVIDIA TensorRT for RTX, delivering up to a 2.5x speedup and a 560% reduction in memory usage. Creators can start generating images locally immediately after unboxing, without separate model downloads or ComfyUI configuration.
NVIDIA AI for Media SDK is adding updates including LipSync NIM microservices tuned for French, German and Spanish, and an enhanced Active Speaker Detection NIM with multicamera support and cross-video speaker correlation.
NVIDIA is directing developers to its latest RTX AI Garage blog post covering Hermes Agent and self-improving AI on RTX PCs and DGX Spark.
The company is promoting RTX Spark and its AI ecosystem across social channels including Facebook, Instagram, TikTok and X, and through an RTX Spark newsletter.
Paul Balo is the founder of TechBooky and a highly skilled wireless communications professional with a strong background in cloud computing, offering extensive experience in designing, implementing, and managing wireless communication systems.
We use cookies to ensure that we give you the best experience on our website. If you continue to use this site we will assume that you are happy with it.
