A laptop screen illustrates an ONNX model flowing through Qualcomm QNN to a Hexagon NPU.
Microsoft has released KB5136904, an update to the Qualcomm QNN Execution Provider for Windows 11, version 26H1. It moves the component to version 2.2611.1.0. You won't see a new Start menu button or a feature splash. This is a small AI plumbing update, and it matters mainly if your Snapdragon PC runs local machine-learning models through Windows ML and ONNX Runtime.

Microsoft's support notice is short. The rest of this article covers what the KB says, what Microsoft's release history adds, and where the documentation stops.

A laptop screen illustrates an ONNX model flowing through Qualcomm QNN to a Hexagon NPU. What KB5136904 actually is​

Microsoft's KB article describes the Qualcomm QNN Execution Provider as an execution provider for ONNX Runtime and for Windows machine-learning scenarios built on ONNX Runtime. It gives hardware acceleration on Qualcomm chipsets. The ONNX Runtime documentation explains how it works: the provider uses the Qualcomm AI Engine Direct SDK (QNN SDK) to construct a QNN graph from an ONNX model which can be executed by a supported accelerator backend library.

Put simply, an app gives Windows ML a standard ONNX model. The QNN provider converts that model into a form Qualcomm's AI hardware can run, then hands it to that hardware. Without the provider, the model typically runs on the CPU, which works but uses more power and is often slower.

The essentials from the KB:

  • Applies to: Windows 11, version 26H1, all editions
  • New version: 2.2611.1.0
  • Delivery: Downloaded and installed automatically through Windows Update
  • Prerequisite: The latest cumulative update for Windows 11, version 26H1
  • Replaces: KB5128782
  • Update history entry: "Windows ML Runtime Qualcomm QNN Execution Provider Update (KB5136904)"

Microsoft's summary says only that the release "includes improvements" to the component. It lists no fixes, performance figures, security changes, newly supported models or known issues. If you see claims about big speed gains in this update, Microsoft's notice doesn't support them.

Section summary: KB5136904 is an automatic component update for Qualcomm's AI acceleration layer on Windows 11 26H1. Microsoft hasn't said what changed.

One of five same-day releases​

Microsoft's release history for AI updates shows this wasn't a Qualcomm-only release. For Windows 11, version 26H1, Microsoft lists five execution provider updates dated November 19, 2026:

Execution providerVersionKB
AMD MIGraphX2.2611.1.0KB5136908
AMD Vitis2.2611.0.0KB5136907
Intel OpenVINO2.2611.1.0KB5136906
Nvidia TensorRT-RTX2.2611.1.0KB5136905
Qualcomm QNN2.2611.1.0KB5136904

The same release history lists a parallel QNN update, version 2.2611.1.0 as KB5136899, for Windows 11 versions 24H2, 25H2 and 26H2, alongside matching AMD, Intel and Nvidia packages.

This matters for anyone managing a mixed fleet. A Snapdragon PC on 26H1 should get KB5136904. A Snapdragon PC on 24H2, 25H2 or 26H2 gets a different KB number for the same version. If your compliance reports only search for "KB5136904," they will undercount on every machine that isn't running 26H1.

The shared version numbers (2.2611.x across vendors) look like a common Windows ML servicing scheme, with "2611" plausibly pointing to November 2026. That's an inference from the pattern, though. Microsoft doesn't explain the numbering in the KB, and it doesn't link 2.2611.1.0 to a specific Qualcomm SDK release.

Section summary: KB5136904 is the 26H1 build of a coordinated November 19 release covering all major AI silicon vendors. Other Windows 11 versions get the same QNN version under a different KB.

How fast this component is moving​

KB5136904 continues a steady release pace for 26H1. WindowsForum's earlier coverage shows the predecessor KB5128782 was released on September 22, 2026, updating the Qualcomm QNN Execution Provider to version 2.2609.3.0 on Windows 11, version 26H1 PCs. Before that, KB5123778 updated the provider to version 2.2608.0.0 on August 27, 2026, replacing KB5121776, which had brought it to version 2.2607.0.0 on July 28, 2026. KB5121776 in turn replaced KB5096137, which updated the provider to version 2.2605.2.0 on supported 26H1 devices, replacing KB5089618, which had brought the component to version 2.2604.2.0.

The 26H1 lineage so far:

  • KB5089618 — 2.2604.2.0
  • KB5096137 — 2.2605.2.0
  • KB5121776 — 2.2607.0.0
  • KB5123778 — 2.2608.0.0
  • KB5128782 — 2.2609.3.0
  • KB5136904 — 2.2611.1.0

That works out to roughly one release every month or two. Microsoft now services the AI runtime more like a driver than an OS feature, with a separate release track that ships updates as Qualcomm's stack changes.

Which hardware actually benefits?​

"All editions" in the KB refers to Windows editions, not to every PC. The QNN provider only does useful work on Qualcomm silicon. Microsoft's Windows ML provider documentation lists these requirements for QNNExecutionProvider: Snapdragon X Elite (X1Exxxxx) or Snapdragon X Plus (X1Pxxxxx), each with a Qualcomm Hexagon NPU on driver version 30.0.140.0 or later.

Some caveats apply:

  • That requirements list comes from Microsoft's general Windows ML provider page. KB5136904 itself doesn't list supported chipsets, so don't read the page as a complete compatibility matrix for this KB.
  • The KB doesn't say it updates the Hexagon NPU driver. Driver updates arrive separately, usually through the OEM or Windows Update driver channels. A current QNN provider paired with an outdated NPU driver can still fail the requirement above.
  • Having a provider doesn't mean the whole model runs on the NPU. Microsoft's Windows ML documentation warns that operator, data-type and tensor-shape support varies, and that graph partitioning and fallback depend on the model and session settings. Parts of a model can still end up on the CPU.

ONNX Runtime's QNN documentation shows the developer side of this. It offers a session option, session.disable_cpu_ep_fallback, which disables fallback of unsupported operators to the CPU EP, and that option is useful when you need to confirm whether a model runs fully on Qualcomm hardware.

Section summary: Snapdragon X-series Copilot+ PCs with a current Hexagon NPU driver are the main beneficiaries. On Intel, AMD and Nvidia systems, the matching vendor KBs in the November 19 release are the relevant ones.

How to verify the update on your PC​

According to Microsoft, you don't need to do anything manually. To confirm the update:

  • Check your Windows version. Go to Settings > System > About and confirm the device runs Windows 11, version 26H1. On 24H2, 25H2 or 26H2, look for the corresponding KB instead (KB5136899 for QNN, per Microsoft's release history).
  • Install the latest 26H1 cumulative update. The KB requires it. If the device is behind on cumulative updates, KB5136904 isn't expected to appear.
  • Run Windows Update. Go to Settings > Windows Update and select Check for updates.
  • Confirm the install. Open Settings > Windows Update > Update history and look for Windows ML Runtime Qualcomm QNN Execution Provider Update (KB5136904).

If it doesn't appear:

  • First check the cumulative update prerequisite. It's the most common reason.
  • Check that the device actually uses a Qualcomm processor, since this is Qualcomm-specific.
  • Check whether the device runs a different Windows 11 version that uses a different KB number.
  • In managed environments, check whether update policies are holding back the latest cumulative update.

The KB gives no standalone download or manual install procedure. Windows Update is the only delivery route it documents.

What developers should know​

Microsoft's Windows ML documentation describes hardware acceleration through Execution Providers (EPs) as the model for local AI inference on Windows. It offers developers two ways to get a provider. Windows-certified providers come through the ExecutionProviderCatalog APIs and update automatically. The "bring your own" option lets you ship provider binaries yourself for offline, managed or version-pinned deployments. That second option is separate developer guidance, not an install method described in KB5136904.

Automatic updates have a tradeoff. You get fixes without shipping anything, but the runtime underneath your app can change between releases. If your app depends on specific operator behavior, run your actual models on real Snapdragon hardware after each provider update. Microsoft's notice lists no changes, so you can't simply check the release notes for problems.

Microsoft's Windows ML guidance also recommends pre-compiling and caching models for a target provider to avoid long cold starts. After a provider update, check whether your compiled cache is still valid.

The bottom line​

KB5136904 is a routine but useful update. Snapdragon X Elite and X Plus PCs on Windows 11 26H1 get Qualcomm QNN Execution Provider 2.2611.1.0 automatically, as long as they have the latest cumulative update. Microsoft hasn't said what changed. For users, the main tasks are installing the cumulative update and checking that the entry shows up in Update history. For admins, the key point is that other Windows 11 versions get the same QNN version as KB5136899, so reports keyed only to KB5136904 will miss them. For developers, it's another reason to retest on real hardware after each update.


Correction (November 20, 2026): An earlier version of this article omitted two releases from the 26H1 Qualcomm QNN lineage: KB5121776 (version 2.2607.0.0, July 28, 2026) and KB5123778 (version 2.2608.0.0, August 27, 2026). The lineage list and history paragraph have been corrected to include them.

 

References

  1. KB5136904: Qualcomm QNN Execution Provider update (2.2611.1.0) | Microsoft Support Microsoft Support 2026-11-19T00:00:00+00:00
  2. KB5136904: Qualcomm QNN Execution Provider update (2.2611.1.0) | Microsoft Support support.microsoft.com
  3. GitHub - onnxruntime/onnxruntime-qnn: onnxruntime-qnn is the Qualcomm AI Runtime (QAIRT) execution provider for onnxruntime. It provides onnxruntime hardware acceleration and advanced functionalities on Qualcomm devices. · GitHub github.com