[Feature Request] Release onnxruntime-openvino v5.10 built against OpenVINO 2026.x — current v5.9 (OV 2025.4) is incompatible with Intel NPU driver >= 32.0.100.4723 (ze_loader 1.28.2)

[laulau]

Summary

onnxruntime-openvino v5.9 (the current latest, ORT 1.24.1 + OpenVINO 2025.4.1) crashes any host application with an unrecoverable access violation (0xc0000005) on machines running Intel NPU driver >= 32.0.100.4723 (ze_loader 1.28.2). These drivers have been distributed via Windows Update since March 2026.

OpenVINO 2026.x does not have this crash. The fix exists. We just need a new onnxruntime-openvino release built against it.

Context

We build a Windows security product (antivirus) using onnxruntime-openvino for ML inference on Intel Core Ultra 9 285K (Arrow Lake). After weeks of investigation (WER crash dumps, PDB stack walking, driver bisection), we identified the full root cause:

BIOS < 3202 (ASUS Z890)
  → NPU driver 4778 installs incorrectly (ACPI table mismatch)
  → ze_loader 1.28.2 initializes with incomplete NPU ACPI data
  → OVDevices::get_ov_devices() → 0xc0000005 in onnxruntime::ProviderLibrary::Load
  → host process killed, no recovery possible

With BIOS 3202 + driver 4778, the NPU driver installs correctly and ze_loader 1.28.2 initializes properly. But OV 2025.4.1 still crashes with ze_loader 1.28.2 — this is an API-level incompatibility that the correct BIOS cannot fix.

OpenVINO 2026.2.0 (Python) on the same machine, same driver, same ze_loader: zero crash.
available_devices = ['CPU', 'GPU.0', 'GPU.1', 'NPU'] — fully working.

Current status for application developers

There is no usable path for onnxruntime-openvino on a fully updated Intel Core Ultra 200 machine today:

BIOS	NPU Driver	ze_loader	OV 2025.4 EP	Works?
< 3202	4778 (latest)	1.28.2	Crash 0xc0000005	No
3202	4778 (latest)	1.28.2	Crash 0xc0000005	No
any	4023 (May 2025)	1.27.x	Works	Yes — but 1 year old

The only working configuration requires a one-year-old driver that is no longer the recommended download on Intel's official page.

We had no choice but to remove OpenVINO entirely from our ML router and fall back to pure ORT CPU inference. Our users lose NPU/iGPU hardware acceleration until v5.10 ships.

Request

Please release onnxruntime-openvino v5.10 built against OpenVINO 2026.x.

The fix exists in OpenVINO's codebase. OV 2026.2 works correctly with ze_loader 1.28.2. Packaging it into a new onnxruntime-openvino release is the only thing standing between a broken state and a working one for all developers targeting Intel Core Ultra 200 on Windows.

A patch release to v5.9 (backporting the ze_loader 1.28.x fix from OV 2026.x to OV 2025.4) would also be acceptable and likely faster to ship.

Timeline pressure

Intel NPU drivers 4723/4778 have been on Windows Update for 3 months. Every developer who updated their machine in this period cannot use onnxruntime-openvino. The longer v5.10 takes, the wider the gap between "machines that can run the EP" and "machines that exist in the wild."

Related: openvinotoolkit/openvino#36456

[n1harika]

@laulau thanks for the detailed investigation.

As a quick workaround, we recommend one of the following options:

1. Migrate to the WinML Execution Provider, which provides an alternative ONNX Runtime execution provider for Intel OpenVINO acceleration on supported systems:

   • Python wheels: https://pypi.org/project/onnxruntime-ep-openvino/
   • NuGet package: https://www.nuget.org/packages/Intel.ML.OnnxRuntime.EP.OpenVINO
     (ONNX Runtime must be installed separately. Usage and documentation are available on the pages above.)

2. Build ONNX Runtime from source against OpenVINO 2026.1 or later, to generate the wheel, where the issue is not observed: https://onnxruntime.ai/docs/build/eps.html#windows-5