KB5128778 moves the 26H1 MIGraphX provider to version 2.2609.2.0
Microsoft's support notice is short. The update "includes improvements to the MIGraphX Execution Provider AI component for Windows 11, version 26H1." It downloads and installs automatically from Windows Update. The only listed prerequisite is the latest cumulative update for Windows 11 version 26H1. KB5128778 replaces KB5103219, which Microsoft lists as MIGraphX Execution Provider version 2.2606.1.0 and which carried the same 26H1 scope and the same automatic delivery. This AMD MIGraphX Execution Provider update includes improvements to the MIGraphX Execution Provider AI component for Windows 11, version 26H1. This update will be downloaded and installed automatically from Windows Update. You must have the latest cumulative update for Windows 11, version 26H1 installed.
Microsoft doesn't list specific fixes, new operators, performance figures or security content. This is a nonsecurity AI component update. Nothing in the notice supports calling it a vulnerability fix or promising that any particular model will run faster.
To confirm installation, open Settings > Windows Update > Update history. A device with the update shows an entry named "Windows ML Runtime AMD MIGraphX Execution Provider (KB5128778)." That is the only verification step Microsoft documents in the KB.
Windows ML services the AMD MIGraphX path separately from the OS
The KB makes more sense once you see how Windows ML is built. An execution provider (EP) is a hardware-specific backend that ONNX Runtime uses to run model operations on a particular chip. Microsoft describes MIGraphX as an AMD execution provider for ONNX Runtime and Windows machine learning. It speeds up inference by moving supported ONNX model operations onto AMD GPUs, and it is built on AMD's MIGraphX graph inference engine. ONNX Runtime's own documentation describes it in similar terms: the MIGraphX execution provider uses AMD's Deep Learning graph optimization engine to accelerate ONNX model on AMD GPUs.
The word "supported" is doing real work there. The provider takes the operations it can handle, and the rest of a model's graph can run elsewhere. The KB says nothing about which operations are covered, so installing it tells you nothing new about whether your particular model will run fully on the GPU.
Windows ML ships only two providers in the box: CPU and DirectML, which Microsoft now labels legacy. Vendor providers are downloaded on demand. Microsoft Learn lists them as AMD MIGraphX (MIGraphXExecutionProvider), AMD VitisAI, NVIDIA NvTensorRtRtx, Intel OpenVINO, Qualcomm QNN and an experimental Microsoft WebGPU provider. Microsoft says these downloadable providers are available on Windows 11 version 24H2 (build 26100) or later, depending on device and driver compatibility, through the Windows ML ExecutionProviderCatalog APIs. That explains why this update ships under its own KB number and isn't folded into a cumulative update: each vendor's provider is serviced as a separate package.
Keep two scopes apart. The general platform documentation covers Windows 11 24H2 and later. KB5128778 is listed only for 26H1.
MIGraphX hardware and driver requirements still decide whether any of this runs
Getting the KB doesn't make the provider usable. Microsoft Learn's general requirements for MIGraphX are:
- An AMD RDNA 3 or later GPU.
- AMD GPU driver version 25.10.13.09 or later.
- No GenAI scenarios. Microsoft says the provider "is not supported for GenAI scenarios today."
These are standing requirements for the provider, not new conditions added by KB5128778. The KB itself asks only for the 26H1 cumulative update. For developers, the GenAI limit is the one that matters most. If your app runs generative models through Windows ML, MIGraphX isn't a supported path whatever version Windows Update delivers. Its supported role is conventional ONNX inference.
AMD also has a second Windows ML provider, VitisAI, which Microsoft's documentation ties to AMD's NPU driver stack. The two are easy to confuse. KB5128778 covers only the GPU-oriented MIGraphX provider.
Monthly provider updates arrive in the D-week slot
Microsoft's Windows ML servicing guidance explains why notices like this appear regularly. Once an execution provider is installed, updates come through Windows Update's optional nonsecurity preview releases, also called "D week" releases. Apps get compatible updates automatically without updating themselves. Microsoft says this lets apps pick up performance improvements and new operator support without code changes. The guidance also states that there is currently no way for an app to programmatically install an updated EP.
That general description explains the mechanism, but it doesn't mean KB5128778 in particular adds operators or improves performance.
Microsoft's KB history shows a steady cadence on two parallel tracks. On Windows 11 24H2 and 25H2, the chain runs from KB5072093 (version 1.8.35.0), which did not replace any previously released update, to KB5077527 at 1.8.43.0, then KB5083461, then KB5089167 at version 2.2604.1.0, which replaced KB5083461, and then KB5096143 at 2.2605.1.0, which replaced KB5089167.
The 26H1 track began separately. KB5078980, at version 1.8.43.0, was the first MIGraphX update for Windows 11, version 26H1 and did not replace any previously released update. Later, KB5096141 delivered version 2.2605.1.0 to 26H1, replacing KB5089171. Then came KB5103219 at 2.2606.1.0, and now KB5128778 at 2.2609.2.0.
| KB | Version | Windows 11 scope | Replaces |
|---|---|---|---|
| KB5078980 | 1.8.43.0 | 26H1 | None |
| KB5096141 | 2.2605.1.0 | 26H1 | KB5089171 |
| KB5103219 | 2.2606.1.0 | 26H1 | KB5096143 (per Microsoft's notice) |
| KB5128778 | 2.2609.2.0 | 26H1 | KB5103219 |
The replacement records contain one inconsistency. Microsoft's KB5103219 notice says "This update replaces the previously released update KB5096143." But KB5096143 is the 2.2605.1.0 package for 24H2 and 25H2, while KB5096141 is the 26H1 package of the same version. This is probably a clerical slip in an older notice, and it doesn't change how KB5128778 behaves. Still, if you track supersedence chains in inventory tools, don't assume the records line up exactly across both tracks.
Two readings of the version numbers are our inference, since Microsoft doesn't document either. The jump from 1.8.x to 2.x lines up with Microsoft Learn's separate provider lists for Windows ML 1.8.x (Microsoft.WindowsAppSDK.ML 1.8.x) and Windows ML 2.x (Microsoft.WindowsAppSDK.ML or Microsoft.Windows.AI.MachineLearning 2.x). The middle field, 2604, 2605, 2606 and now 2609, looks like a year-and-month stamp. The move from 2606 to 2609 also suggests there was no 26H1 MIGraphX release in July or August, or at least none that Microsoft replaced in this chain. The "2" in the last position of 2.2609.2.0, where earlier releases ended in ".1.0", may point to a respin within the month, but Microsoft hasn't said so.
What KB5128778 means for AMD GPU users and developers
For most readers there's nothing to decide: the update installs itself on eligible 26H1 machines. The steps below matter if you manage 26H1 devices with RDNA 3 or later Radeon GPUs, or if you ship Windows ML apps that rely on MIGraphX.
If you manage 26H1 PCs, make sure the latest 26H1 cumulative update is installed, because KB5128778 depends on it. Then check Update history for the "Windows ML Runtime AMD MIGraphX Execution Provider (KB5128778)" entry. Devices on 24H2 or 25H2 won't get this KB, and its absence there isn't a fault.
If you develop Windows ML apps, remember that the provider under your app can change once a month without you shipping anything. Microsoft pitches this as the benefit of the model, since operator and performance improvements arrive for free. It also means inference behavior on a user's machine can shift between releases of your app. Because apps can't install EP updates programmatically, the practical step is to record the EP version a device has when you're triaging bug reports. Microsoft's Windows ML documentation includes a page on checking EP versions for that purpose. Also plan for the GenAI exclusion: don't send generative workloads to MIGraphX and expect support.
If you're troubleshooting a machine where MIGraphX never activates, start with the platform requirements rather than the KB. You need an RDNA 3 or newer GPU and AMD driver 25.10.13.09 or later. A device that fails those checks won't use MIGraphX however current its KB is.
- KB5128778 updates the AMD MIGraphX Execution Provider to version 2.2609.2.0 on Windows 11 version 26H1 only, all editions, and replaces KB5103219 (version 2.2606.1.0).
- Delivery is automatic through Windows Update, and the latest Windows 11 26H1 cumulative update must already be installed.
- Microsoft lists only unspecified "improvements," with no security fixes, benchmarks or new-operator details, so don't plan around a specific gain.
- To confirm installation, look under Settings > Windows Update > Update history for "Windows ML Runtime AMD MIGraphX Execution Provider (KB5128778)."
- MIGraphX still needs an AMD RDNA 3 or later GPU and driver 25.10.13.09 or later, and Microsoft doesn't support it for GenAI scenarios.
- Microsoft's replacement records for the older KB5103219 point to a 24H2/25H2 package, so check supersedence chains before relying on them in inventory tooling.
KB5128778 is a routine step in a monthly cadence. Microsoft now services each silicon vendor's AI backend separately from the OS, and gives each release its own KB and Update history entry. The practical work is to confirm the prerequisite cumulative update, check the history entry, and log EP versions when diagnosing inference problems. Going by the version stamps, the next MIGraphX package for 26H1 should replace KB5128778 in a later optional D-week release, and it will most likely come with the same one-line summary.