Nvidia introduced RTX Spark at GTC Taipei during COMPUTEX, a new class of Windows PC purpose-built to run personal AI agents on-device, packing 1 petaflop of AI compute and 128GB of unified memory. The machine ships this fall alongside a software stack that includes Nvidia's OpenShell runtime for Windows and a refreshed NemoClaw blueprint spanning GeForce RTX, RTX PRO, DGX Spark and DGX Station. Nvidia also announced DGX Station for Windows, a deskside system pairing a data-center-class GPU and CPU with Windows for manageability and security.
The pitch is straightforward: agents that can read files, drive applications and chain multi-step tasks need both privacy guarantees and serious local compute. RTX Spark targets that workload directly, positioning the hardware against cloud inference for users unwilling to send personal context off-device. Nvidia frames the launch as drawing on 30 years of technology innovation, with form factors ranging from slim laptops with all-day battery life to ultraefficient desktops.
“RTX Spark's 1 petaflop of AI compute and 128GB of unified memory can meet the computing demand of on-device agents, offering a new class of computer that goes from tool to teammate.”— Gerardo Delgado, Nvidia
On the software side, Nvidia and Microsoft are co-developing security primitives for Windows agents that handle identity, containment and policy enforcement. OpenShell layers on top, letting users define what agents can do, route queries to local models based on privacy rules, and strip personal information from prompts headed to cloud models. Hermes Agent and OpenClaw, two open-source agent projects with rapid GitHub uptake, are adopting OpenShell and the Microsoft primitives in their new Windows apps.
Key facts
- 01RTX Spark ships this fall with 1 petaflop of AI compute and 128GB of unified memory, aimed at running personal AI agents locally on Windows.
- 02Optimizations to llama.cpp deliver 2x throughput on Qwen3.6-27B and a 1.6x boost on Qwen3.6-35B via multi-token prediction.
- 03Nvidia's OpenShell runtime is coming to Windows, built on Microsoft's new security primitives for on-device agents.
- 04Quantized Holo Computer Use models from H Company run 2x faster on Nvidia GPUs while cutting memory consumption by 35%.
- 05Adobe is rearchitecting Photoshop and Premiere for RTX Spark, promising 2x faster AI, editing, coloring and effects.
Performance work on open models is the other half of the story. Nvidia collaborated with the llama.cpp community on multi-token prediction, a speculative decoding technique that delivers 2x throughput on Qwen3.6-27B and a 1.6x boost on Qwen3.6-35B when measured on a GeForce RTX 5090. For multi-GPU rigs, llama.cpp adds tensor parallelism for up to 2x memory and 1.8x compute on two equivalent GPUs, while ComfyUI gains a new classifier-free guidance method worth up to 2x performance on the same configuration.
DGX Spark, Nvidia's Linux-side personal agent computer, gets a parallel set of updates. A streamlined NemoClaw installer arrives this month with automatic sandboxing and Hermes Agent support, available across RTX and DGX PCs on Linux and Windows Subsystem for Linux. Working with the vLLM project, Nvidia shipped new NVFP4 checkpoints for Qwen3.6-35B that deliver 2.6x performance on DGX Spark compared with the previously available checkpoints from Unsloth.
Nvidia is also pushing into computer-use agents through a partnership with H Company. Its Holo Computer Use models are being quantized for Nvidia GPUs, with the company reporting a 2x speedup and a 35% reduction in memory consumption. The Holo Desktop app is due soon and will let agents drive a PC by seeing the screen and using the mouse and keyboard, even in applications with no API surface.
“H Company's computer-use harness lets agents navigate a PC by seeing the screen and operating a mouse and keyboard just like a user, even in apps with no application programming interfaces.”— Gerardo Delgado, Nvidia
Adobe is rebuilding Photoshop and Premiere around RTX Spark's hardware. The reworked stack promises up to 2x faster AI, editing, coloring and effects, with Premiere getting a new video pipeline that taps unified memory, the Blackwell GPU and TensorRT for real-time editing and color correction. Photoshop's next-generation engine targets GPU-accelerated compositing for live filters, HDR and natural brushing, while Substance 3D Painter and Stager will run natively on RTX Spark. Adobe also plans to expose Premiere and Photoshop to Windows agents so users can edit through agent workflows. The updates start rolling out alongside RTX Spark this fall.
Smaller updates round out the announcement. Nvidia Broadcast 2.2 graduates Studio Voice out of beta starting today, now running on GeForce RTX 3060 GPUs and above, and adds Elgato Stream Deck support. Project G-Assist also picks up Stream Deck integration, extending agent and creator hooks to hardware controllers.
The open questions are price, battery life on shipping laptops, and how cleanly OpenShell's privacy routing works in practice. Nvidia did not disclose RTX Spark pricing, and the security model leans heavily on Microsoft primitives that have not yet been independently audited at scale. The 2x and 1.6x numbers on Qwen models are real but reflect best-case configurations with the latest optimizations; agent reliability on long-horizon, cross-app tasks remains the harder problem the industry has not yet solved.
For the broader AI market, the move matters because it stakes Nvidia's claim on the local-agent layer just as Microsoft pushes Copilot+ PCs and Apple readies on-device Apple Intelligence. The 128GB unified memory figure is the headline spec — it puts mid-size open models within reach of a desktop without cloud round-trips, which is the actual bottleneck for agent latency. If Nvidia can convert RTX Spark into a credible developer platform with Adobe, Hermes and OpenClaw shipping on day one, the company turns its GPU lead into a software flywheel that the cloud-first agent stacks will have to answer.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
