A laptop displays a software update panel as glowing circuit lines connect it to a processor and download icon.
Microsoft released KB5128782 on September 22, 2026. It updates the Qualcomm QNN Execution Provider, the Windows AI component that runs ONNX machine-learning models on Qualcomm hardware, to version 2.2609.3.0 on Windows 11, version 26H1 PCs. Windows Update installs it automatically once the latest 26H1 cumulative update is on the device. This is a routine component update, and Microsoft's notice says only that it contains "improvements." It lists no specific fixes, performance gains or new model support. For most owners of Snapdragon-based 26H1 machines there is nothing to do. For administrators and developers, the main value is knowing which KB belongs to which Windows release and how to confirm it installed.

KB5128782 moves the QNN Execution Provider to 2.2609.3.0 on Windows 11 26H1​

Microsoft's support notice covers Windows 11, version 26H1 in all editions. It describes the update as improvements to "the Qualcomm QNN Execution Provider AI component." Delivery is automatic through Windows Update. The one stated prerequisite is the latest cumulative update for Windows 11, version 26H1. KB5128782 replaces KB5123778, the previous QNN provider release.

Two identifiers matter here. KB5128782 is the Windows servicing package. 2.2609.3.0 is the version of the QNN provider inside it. Update history shows the KB number, not the component version. After installation, the entry should read "Windows ML Runtime Qualcomm QNN Execution Provider Update (KB5128782)."

The notice leaves out a lot. It gives no changelog, known issues, package size, restart behaviour, manual-download option or deployment deadline. It also names no Windows feature or app that benefits. Any claim that 2.2609.3.0 makes a particular AI feature faster or fixes a particular bug goes beyond what Microsoft has published.

How the Qualcomm QNN Execution Provider turns ONNX models into NPU work​

An execution provider is a plug-in for ONNX Runtime, the inference engine that Windows ML is built on. It connects a model to a specific type of hardware. In Microsoft's general description of these components, an execution provider handles graph partitioning, kernel selection and operator execution. It also hides vendor-specific acceleration libraries, so one model can run on different hardware without changes to the app. Microsoft says these components are used on Copilot+ PCs by Windows features, Copilot+ experiences and third-party apps that run AI models locally. When hardware acceleration isn't supported for a workload, execution falls back to another compatible backend.

The QNN provider is Qualcomm's version of this. It uses the Qualcomm AI Engine Direct SDK (QNN SDK) to construct a QNN graph from an ONNX model, which is then executed by a supported accelerator backend library. In practice the provider translates a standard model format into a form Qualcomm's own runtime can schedule on the chip's accelerators.

The upstream ONNX Runtime documentation adds a detail that matters for developers. The QNN HTP backend only supports quantized models. Models with 32-bit floating-point activations and weights must first be quantized to use a lower integer precision (e.g., 8-bit or 16-bit integers). HTP is the backend that targets the Hexagon NPU. This is general ONNX Runtime behaviour, not something KB5128782 changes. It does explain why providers like QNN get serviced so often: they sit at a sensitive layer between models and silicon.

Microsoft Learn's Windows ML documentation lists the provider's registered name as QNNExecutionProvider. The requirements it gives are a Snapdragon X Elite (X1E-series) or Snapdragon X Plus (X1P-series) chip with a Qualcomm Hexagon NPU driver at version 30.0.140.0 or later. Those are general Windows ML requirements for the provider. KB5128782 doesn't restate them as prerequisites, and they shouldn't be read as a complete list of devices that will receive this update.

Why a Windows ML component ships outside the cumulative update​

Microsoft Learn separates execution providers that ship with Windows ML from those that don't. The CPU provider and the legacy DirectML provider are built into the ONNX Runtime that ships with Windows ML. Vendor providers are not. QNN, Intel's OpenVINO, NVIDIA's TensorRT-RTX, and AMD's MIGraphX and Vitis AI are all downloaded on demand. On Windows 11 24H2 (build 26100) and later, developers can obtain them through the Windows ML ExecutionProviderCatalog APIs, depending on device and driver compatibility.

That split helps explain why QNN has its own KB numbers instead of riding inside Patch Tuesday updates. The provider is a separately versioned component, and Microsoft services it that way. The KB5128782 notice only says delivery happens through Windows Update. It doesn't say whether that path shares plumbing with the catalog APIs developers use, so treat the two as related but not proven identical.

The cumulative-update prerequisite follows the same logic. The provider update assumes a current base OS. A 26H1 machine that has fallen behind on cumulative updates shouldn't be expected to pick up KB5128782 until it catches up.

The 26H1 track and the 24H2/25H2 track follow separate KB chains​

Microsoft's QNN notices show two parallel servicing lines, and the 26H1 one has moved quickly. The first 26H1 release on record is KB5078978, version 1.8.30.0, and this update does not replace any previously released update. From there the chain runs through several steps:

KBQNN provider versionWindows 11 releaseReplaces
KB50789781.8.30.026H1Nothing
KB50961372.2605.2.026H1KB5089618
KB51217762.2607.0.026H1KB5096137
KB51237782.2608.0.026H1KB5121776
KB51287822.2609.3.026H1KB5123778

Microsoft's KB5096137 notice confirms it is for 26H1 and that this update replaces the previously released update KB5089618. The KB5121776 notice gives the next step: this update replaces the previously released update KB5096137. Microsoft's KB5123778 notice lists version 2.2608.0.0, 26H1 only, and names KB5121776 as its predecessor. That makes KB5128782 the latest link in an unbroken chain.

The 24H2/25H2 line uses different KB numbers. KB5096135 carries the same 2.2605.2.0 provider version and includes improvements to the Qualcomm QNN Execution Provider AI component for Windows 11, version 24H2 and Windows 11, version 25H2. This update replaces the previously released update KB5089617, which in turn followed KB5077526. So an identical provider version can ship under different KB numbers depending on the Windows release.

For administrators the practical point is simple. KB5128782 is a 26H1 package. Don't expect it on 24H2 or 25H2 Snapdragon PCs, whose QNN updates come under their own KB numbers. Nothing in the September 22 notice says whether a matching 2.2609.3.0 release exists for 24H2/25H2.

A note on the numbers themselves: since the 2.x series began, the middle four digits (2605, 2607, 2608, 2609) appear to follow a year-and-month pattern. That's an observation from the version strings, not something Microsoft has documented, so it shouldn't be used to predict release dates.

Confirming KB5128782 on a 26H1 Snapdragon PC​

The only verification step Microsoft documents is the Settings app. You can't use it to force the update.

  1. Make sure the device is on the latest cumulative update for Windows 11, version 26H1. This is the stated prerequisite.
  2. Let Windows Update run normally. KB5128782 downloads and installs automatically.
  3. Open Settings > Windows Update > Update history.
  4. Look for "Windows ML Runtime Qualcomm QNN Execution Provider Update (KB5128782)." If you see it, the 2.2609.3.0 provider is installed.

If the entry is missing, first check the two stated boundaries: is the machine actually on 26H1, and is it fully up to date on cumulative updates? Beyond that, Microsoft offers no troubleshooting steps, no standalone installer and no rollback procedure. A missing entry on a machine outside those boundaries tells you nothing about whether it is eligible. On a machine inside them, it means the update may simply not have arrived yet.

Seeing KB5123778 in the history instead means the device still has the 2.2608.0.0 provider. That's the older release that KB5128782 replaces.

What this means for you​

For almost everyone the decision is to let the update install automatically and move on. Only administrators who track exact component versions, and developers testing ONNX models on Qualcomm NPUs, have a reason to confirm which QNN release a machine has.

Developers should keep in mind that the notice lists no changed operators, APIs, supported data types or model-compatibility changes. If an ONNX model behaves differently on a 26H1 Snapdragon test machine after this update, compare that machine's Update history against KB5123778 and KB5128782 to narrow down whether the provider changed. Don't assume a regression or an improvement without testing it yourself.

  • KB5128782 installs Qualcomm QNN Execution Provider 2.2609.3.0 on Windows 11, version 26H1 and replaces KB5123778, which carried version 2.2608.0.0.
  • Windows Update installs it automatically, but only on devices that have the latest Windows 11 26H1 cumulative update.
  • Update history confirms installation with the entry "Windows ML Runtime Qualcomm QNN Execution Provider Update (KB5128782)."
  • Windows 11 24H2 and 25H2 PCs receive QNN updates under different KB numbers, so KB5128782 won't appear on them.
  • Microsoft describes the change only as "improvements," with no documented fixes, performance figures or known issues.
  • Microsoft Learn's general requirements for the QNN provider are a Snapdragon X Elite or X Plus chip with a Hexagon NPU driver at 30.0.140.0 or later. Check those before blaming the provider for missing NPU acceleration.

This is the fifth QNN provider release on the 26H1 track, and the 2.x versions have come in quick succession. That pace shows Microsoft treats NPU execution providers as fast-moving components serviced alongside the OS rather than built into it. Nothing here needs a response unless you're a 26H1 administrator or a developer checking which provider version a test machine runs. For everyone else, Windows Update will replace 2.2609.3.0 with whatever comes next without any action needed.