Microsoft has quietly pushed out a servicing update that upgrades the NVIDIA TensorRT-RTX Execution Provider to version 2.2604.1.0, promising faster local AI inference on Windows 11—but there’s a catch. KB5089174 only applies to Windows 11 version 26H1, a specialized release limited to select next-generation hardware. For the few devices that qualify, the update means snappier AI workloads without a single visible change to the user interface. For everyone else, it’s a look behind the curtain at how Windows is stitching together a platform for on-device machine learning.
A Silently Installed Component Upgrade
KB5089174 is not a cumulative security roll-up, not a feature update, and certainly not a new NVIDIA graphics driver. It is a targeted refresh of the Windows ML Runtime NVIDIA TensorRT-RTX Execution Provider—a component that helps Windows and third-party applications run local AI models on NVIDIA RTX GPUs. The update bumps the component from version 2.2603.1.0 (delivered in March via KB5083464) to 2.2604.1.0. The installation happens automatically through Windows Update, provided the PC is running Windows 11 version 26H1 and has already ingested the latest cumulative update for that version.
The change is entirely under the hood. Microsoft’s support article describes it as “improvements to the execution provider component.” No bug fixes are listed, no new APIs are documented, and no interface elements are touched. If the update lands successfully, you’ll only know because Windows Update history will display: “Windows ML Runtime Nvidia TensorRT-RTX Execution Provider Update (KB5089174).”
What It Means for You—Home Users, Admins, and Developers
The impact depends heavily on who you are and what hardware you’re running.
Home Users
If your PC is blessed with Windows 11 version 26H1—a release currently reserved for cutting-edge systems shipping with specific hardware innovations—and you have an NVIDIA RTX GPU, this update is a free performance uplift for any AI inference that leans on ONNX Runtime. Applications like local photo upscalers, speech-to-text engines, or creative tools that run machine learning models on-device may see lower latency and better throughput. The difference won’t light up your Start menu, but it could make an AI-powered background blur in a video call feel just a fraction more responsive.
If you’re on Windows 11 23H2, 24H2, or the future 25H2, this KB won’t appear. Even if you have a beast of an RTX 4090, the update is not for you—yet. Microsoft has not signaled plans to backport it to other Windows versions.
IT Administrators
For sysadmins managing fleets of devices on Windows 11 26H1, KB5089174 is a component update that must be tracked separately from monthly cumulative patches. It shows up in its own section of update history and won’t be bundled into the typical “Security Update” bucket. If your compliance reports hinge on detecting KB5083464, they should now look for KB5089174 as the superseding package. The update arrives automatically, but its appearance can be delayed by Windows Update for Business policies, WSUS approval flows, or Intune deferral windows. Ensure your target devices have first installed the latest cumulative update for 26H1; without that prerequisite, this execution provider update won’t deploy.
Crucially, KB5089174 is not a substitute for NVIDIA display drivers. Your fleet still needs regular driver updates through your standard channel—this update only touches the ONNX inference stack, not the graphics pipeline or general GPU compute capabilities.
Developers
If you build applications that leverage Windows ML or ONNX Runtime for local AI, KB5089174 is a platform refinement worth noting. The updated TensorRT-RTX Execution Provider can now build more efficient inference engines on RTX GPUs, potentially reducing model execution time for operations it supports. But your app must be designed to fall back gracefully: the execution provider is only available when the hardware, driver, and Windows version align. On a Windows 11 26H1 device with a compatible RTX GPU, your application may see a performance bump. On any other configuration, the same ONNX graph should still run on the CPU or another backend. Thorough testing across the matrix of Windows versions and GPU models remains essential.
How We Got Here
This isn’t a one-off. Over the past few years, Microsoft has been quietly building a modular AI runtime that lets Windows offload machine learning to the best available accelerator. At the center sits ONNX Runtime, an open-source engine that speaks the ONNX model format. Around it, a pluggable system of execution providers routes specific model ops to hardware-optimized libraries: DirectML for integrated GPUs, Intel OpenVINO for supported Intel silicon, Qualcomm’s QNN for Snapdragon NPUs, and TensorRT-RTX for NVIDIA’s client GPUs.
The TensorRT-RTX execution provider leverages NVIDIA’s TensorRT for RTX runtime, which specializes in converting ONNX graphs into custom inference kernels fine-tuned for the GPU architecture in the machine—think Ampere, Ada Lovelace, or whatever comes next. The result is lower latency and higher throughput than running the same model through a generic CUDA or DirectML path. But that magic only works if the stars align: the model must contain operations the execution provider can accelerate, the GPU driver must be recent enough, and the Windows servicing stack must deliver the updated component.
Windows 11 version 26H1 is unique in this story. It’s not the next mainstream feature update. Instead, Microsoft bills it as a release “designed to support next-generation hardware innovation with device manufacturers and silicon partners.” That likely means systems with novel chip configurations—perhaps new neural processing units, hybrid architectures, or AI-accelerated system-on-chips. Because these devices are expected to run local AI workloads aggressively, keeping the RTX GPU acceleration path up to date makes sense. The previous execution provider version (2.2603.1.0, KB5083464) arrived in the March servicing cycle. This April refresh is a fast-follow iteration, tightening the integration.
What to Do Now
If you suspect you’re on Windows 11 26H1:
- Confirm your Windows version. Open Settings > System > About, or run
winver. The version should say “Windows 11, version 26H1.” - Install the latest cumulative update. Go to Settings > Windows Update and click Check for updates. Install everything pending and restart if prompted.
- After the restart, check update history (Settings > Windows Update > Update history). Look for “Windows ML Runtime Nvidia TensorRT-RTX Execution Provider Update (KB5089174).” If it’s listed, you’re current.
- If the update doesn’t appear, wait a few days. Microsoft’s automatic rollout can be staggered. On managed PCs, consult your IT department—group policies may defer the update.
If you’re on any other Windows 11 version: Do nothing. KB5089174 is not applicable, and attempting to force-install it from the Microsoft Update Catalog could break servicing stack logic.
For developers: Refresh your test environment to Windows 11 26H1 with the latest cumulative and KB5089174. Profile ONNX Runtime sessions with the TensorRT-RTX execution provider enabled. Verify that your app’s fallback to CPU or DirectML still functions correctly on systems without this update.
Why This Matters Beyond the Update
KB5089174 won’t make headlines next to spicy new Copilot features or a major Windows 12 rumor. But it’s a brick in the foundation of an AI-native PC. As more applications bake in local inference—real-time translation, AI-enhanced search, background generation—the difference between a generic ONNX path and a hardware-optimized one becomes the difference between a stutter and a seamless experience. Updates like this keep that plumbing flowing. And while 26H1’s exclusivity limits the audience today, the same execution provider code will eventually ship in the mainstream Windows build when the underlying hardware innovations become the norm.
For now, if your device is one of the few riding the 26H1 wave, let KB5089174 install quietly. You probably won’t notice it—until the next time an AI task completes an instant quicker than you expected.