A Windows 11 workstation displays a local AI inference workflow using ONNX and NVIDIA TensorRT-RTX beside a GeForce PC.
Microsoft has released another monthly update for one of the Windows AI components that most people never see. KB5136905 updates the NVIDIA TensorRT-RTX Execution Provider to version 2.2611.1.0 on Windows 11, version 26H1. It comes through Windows Update, needs no action from most users, and Microsoft gives very little detail about what changed. Here is what the update covers, who gets it, and how to confirm it installed.

A Windows 11 workstation displays a local AI inference workflow using ONNX and NVIDIA TensorRT-RTX beside a GeForce PC. What KB5136905 actually is​

Microsoft's support article says KB5136905 applies to Windows 11, version 26H1, all editions. The update contains "improvements" to the NVIDIA TensorRT-RTX Execution Provider component, and that is all Microsoft says about the contents.

Microsoft describes the component as an ONNX Runtime and Windows machine-learning execution provider that speeds up ONNX model inference on NVIDIA RTX GPUs in consumer PC scenarios. It uses NVIDIA's TensorRT for RTX runtime to build and run RTX-optimized inference engines directly on the local GPU. Windows and apps can then use the RTX card for AI work instead of relying on generic GPU paths or the CPU.

NVIDIA's own documentation explains the design. The provider is built as a standalone plugin that implements the ORT Plugin EP ABI and leverages TensorRT for RTX to accelerate ONNX models on RTX GPUs (Ampere and later).

Section summary: KB5136905 is a version bump for the NVIDIA GPU inference backend in Windows ML on 26H1. It is not a feature update, a driver, or a security patch.

The quick facts​

ItemDetail
KB numberKB5136905
ComponentNVIDIA TensorRT-RTX Execution Provider
New version2.2611.1.0
Applies toWindows 11, version 26H1, all editions
DeliveryAutomatic, via Windows Update
PrerequisiteLatest cumulative update for Windows 11, version 26H1
ReplacesKB5132815
Update History entryWindows ML Runtime NVIDIA TensorRT-RTX Execution Provider Update (KB5136905)

Part of a coordinated November release​

KB5136905 shipped alongside updates for other vendors' execution providers. Microsoft's AI-update release history for Windows 11 version 26H1 lists all of them on the same date. The November 19, 2026 row shows Version2.2611.1.0KB5136908 | Version2.2611.0.0KB5136907 | Version2.2611.1.0KB5136906 | Version2.2611.1.0KB5136905 | Version2.2611.1.0KB5136904, under the AMD MIGraphX, AMD Vitis, Intel OpenVINO, NVIDIA TensorRT-RTX and Qualcomm QNN columns.

Windows 11 versions 24H2, 25H2 and 26H2 got their own NVIDIA package on the same day. It has the same version number but a different KB: Version2.2611.1.0KB5136900 sits in the NVIDIA column for that group.

If you are searching your environment for KB5136905 on a 24H2, 25H2 or 26H2 machine, you will not find it. Those devices get KB5136900 instead. Admins writing compliance reports should map KB numbers to the specific OS release before treating a missing KB as a failed install.

The version history​

KB5136905 replaces KB5132815. Microsoft's release history lists KB5132815 as the October 28, 2026 NVIDIA provider release for 26H1, at version 2.2610.3.0. The version numbers appear to use a year-and-month pattern (2.2610.x in October, 2.2611.x in November). Microsoft does not say what changed between those two builds.

Section summary: This is one of five vendor provider updates released together on November 19. 26H1 devices get KB5136905, and 24H2, 25H2 and 26H2 devices get KB5136900 at the same version.

Who actually benefits?​

Having the update installed does not mean the provider is being used. Several conditions apply:

  • Hardware: Microsoft Learn lists the NVIDIA provider's requirement as GeForce RTX 30-series or newer, with a minimum recommended driver version of 32.0.15.5585 and CUDA 12.5. NVIDIA also says the provider supports RTX GPUs based on Ampere and later architectures (NVIDIA GeForce RTX 30xx and above). These are Windows ML provider requirements. KB5136905 itself lists only the latest 26H1 cumulative update as a prerequisite.
  • App support: An app benefits only if it runs ONNX models through Windows ML and selects this provider. If no installed app uses the provider, the update changes nothing you will notice.
  • Model support: Microsoft says developers are still responsible for optimizing their models for each hardware target. Windows ML handles distribution of execution providers, not model tuning.

Microsoft has published no benchmark, known-issues list, or changelog for KB5136905. Any claim that it makes your games or AI apps faster is unsupported.

How to verify the install​

  1. Make sure the device runs Windows 11, version 26H1 and has the latest cumulative update installed.
  2. Open Settings > Windows Update > Update history.
  3. Look for Windows ML Runtime NVIDIA TensorRT-RTX Execution Provider Update (KB5136905).

Microsoft does not provide a manual download link, a rollout schedule, or troubleshooting steps for this KB. If the entry is missing, the two documented things to check are the OS version and whether the latest cumulative update is installed.

For developers​

Microsoft's Windows ML versioning documentation describes two optional checks:

  • On your own machine: run Get-AppxPackage MicrosoftCorporationII.WinML.* in PowerShell to list installed Windows ML provider packages and their versions.
  • On user devices: read the PackageId property of each provider returned by the ExecutionProviderCatalog APIs. If the provider is not installed, PackageId returns null.

These checks matter because the provider updates automatically. Microsoft notes that most providers are acquired dynamically and receive compatible updates over time, so the version on a user's PC can change between app launches without any action from them.

Why execution provider updates matter​

Windows ML picks a vendor-certified backend for each type of hardware: NVIDIA's provider for RTX GPUs, Intel's OpenVINO, Qualcomm's QNN, and AMD's providers. Microsoft distributes these through Windows Update. An app developer can target one Windows API and leave it to Windows to keep each vendor's backend current.

NVIDIA has argued for this approach since launch. An Edge AI and Vision Alliance post on NVIDIA's work said that on RTX GPUs, Windows ML could run ONNX models with low-latency inference and 50% faster throughput compared to prior DirectML implementations on NVIDIA RTX GPUs. That figure is vendor-reported, dates from the original launch, and has nothing to do with KB5136905 specifically.

There are also trade-offs to consider:

  • Version drift: Automatic updates mean an app can run on a slightly different backend than the one it was tested against. That is why Microsoft offers a "bring your own EP" option for offline, managed or version-pinned deployments.
  • Cold starts: TensorRT for RTX compiles engines just-in-time for the specific GPU. Microsoft recommends precompiling and caching models because a cold start can take seconds or even minutes. That advice is general Windows ML guidance, not something new in this update. NVIDIA's ONNX Runtime documentation says that setting a runtime cache path enables faster model loading by caching JIT compiled kernels for each TensorRT RTX engine.
  • Upstream changes: ONNX Runtime's documentation says the built-in TensorRT RTX Execution Provider in the ONNX Runtime repository is deprecated. We strongly recommend using the standalone EP ABI plugin instead. Developers who use their own ONNX Runtime builds should take note.

NVIDIA's TensorRT for RTX download page shows the kind of work happening upstream. It states that Qwen3.5 dense models are supported through Windows ML with the TensorRT-RTX execution provider. The page does not link that support to version 2.2611.1.0, so it should not be read as a KB5136905 change.

The bottom line​

KB5136905 is a routine monthly refresh. It moves the NVIDIA TensorRT-RTX provider on Windows 11 26H1 to version 2.2611.1.0 and replaces October's KB5132815.

  • Home users: No action needed. The update installs automatically through Windows Update.
  • Admins: Map KB numbers to OS versions. KB5136905 is for 26H1, and KB5136900 is the equivalent for 24H2, 25H2 and 26H2.
  • Developers: Check provider versions through PackageId or PowerShell, precompile and cache models, and test against new provider builds as they arrive.

Microsoft hasn't said what improved in this release. If you run RTX-accelerated AI apps and notice behavior changes after this update, report them in the forums.

 

References

  1. KB5136905: Nvidia TensorRT-RTX Execution Provider update (version 2.2611.1.0) | Microsoft Support Microsoft Support 2026-11-19T00:00:00+00:00
  2. KB5136905: Nvidia TensorRT-RTX Execution Provider update (version 2.2611.1.0) | Microsoft Support support.microsoft.com
  3. GitHub - NVIDIA/TensorRT-RTX-EP-ABI: The NVIDIA TensorRT RTX Execution Provider (EP) is an inference deployment solution designed specifically for NVIDIA RTX GPUs, optimized for client-centric use cases. · GitHub github.com