KB5067994 QNN EP 1.8.13.0 for Copilot+ on Windows 11 24H2

Microsoft has quietly pushed KB5067994, an AI component update that advances the Qualcomm QNN Execution Provider used by ONNX Runtime to version 1.8.13.0 for Copilot+ PCs running Windows 11, version 24H2 — a targeted, per‑silicon runtime refresh delivered automatically through Windows Update that promises incremental improvements to hardware‑accelerated on‑device AI on Qualcomm Snapdragon platforms.

Background / Overview​

The QNN Execution Provider (QNN EP) is the ONNX Runtime plugin that translates ONNX model operators into QNN graph constructs and leverages the Qualcomm AI Engine Direct SDK (QNN SDK / QAIRT) to run those graphs on Qualcomm accelerators (Hexagon NPU/HTP, or GPU/CPU fallbacks). ONNX Runtime documentation describes the provider, its session options (for example, ep.context_enable and disable_cpu_ep_fallback), and the requirement that QNN SDK libraries be present when building from source — though prebuilt ONNX Runtime packages include QNN dependencies starting with certain releases.
Microsoft’s KB for KB5067994 is deliberately succinct: it states the package advances the QNN Execution Provider to version 1.8.13.0, applies to Copilot+ PCs on Windows 11 24H2, requires the latest cumulative update for 24H2 as a prerequisite, and will be installed automatically via Windows Update; it does not present a granular engineering changelog.
This update fits a broader pattern: Microsoft is shipping small, vendor‑targeted AI component updates (Intel OpenVINO EP, NVIDIA TensorRT EP, AMD/Vitis EP and Qualcomm QNN EP) as modular runtime pieces that can be iterated on independently of major OS feature releases. That approach accelerates fixes and per‑silicon tuning but increases operational complexity for administrators who must validate cross‑stack compatibility.

---

What KB5067994 actually does (and what it does not)​

The explicit, verifiable facts​

• Applies only to Copilot+ PCs running Windows 11, version 24H2; it will be installed automatically through Windows Update once the latest cumulative update (LCU) for 24H2 is present.
• The component version installed by this KB is QNN Execution Provider version 1.8.13.0.
• The KB explicitly notes it “includes improvements” but does not list per‑operator or per‑bug fixes, nor does it publish CVE identifiers or a detailed change log. Treat specific performance or security claims as unverified unless Microsoft or Qualcomm disclose them.

What is left unspecified​

Microsoft’s public KB entry intentionally omits a line‑by‑line changelog, leaving several practical questions unanswered unless vendors publish further notes or engineers perform empirical testing:

• Which operators or patterns gained coverage or stability fixes?
• Whether the package addresses any named security issues (no CVE list in the KB).
• Which Qualcomm runtime (QAIRT / QNN SDK) library versions, if any, are required or bundled for specific devices.
• Exact performance deltas on different Snapdragon families (X Elite vs. prior Snapdragon SoCs).

These limitations mean the most reliable verification of tangible benefits is empirical telemetry and controlled testing on representative hardware.

---

Technical context: why QNN EP matters for Copilot+ features​

The QNN Execution Provider is not an isolated application library — it is the bridge between ONNX Runtime models and Qualcomm’s NPU/GPU/CPU execution backends. On Copilot+ devices Microsoft targets with high‑end NPUs, this runtime layer shapes latency, thermal profile, power consumption, and the reliability of local AI workloads such as:

• Short‑form local generative flows (time‑to‑first‑token for system SLMs or client LLM offloads).
• Image processing primitives: segmentation, denoising, super‑resolution used by Photos, Paint, and Studio Effects.
• Camera and video effects in conferencing (virtual backgrounds, live segmentation) and Windows Hello capture paths.

Qualcomm’s AI Hub and QAIRT release notes confirm frequent updates to the QNN runtime, ONNX Runtime integration, and tooling for compiling context binaries and device‑specific artifacts — indicating that incremental provider updates are a normal part of maintaining operator coverage and compilation stability. These vendor materials show the ecosystem’s reality: the QNN stack evolves rapidly and Microsoft’s OS‑level delivery of QNN EP enables on‑device updates without waiting for a large OS feature drop.

---

Likely contents of the 1.8.13.0 bump (reasoned analysis)​

Because the KB itself is terse, a pragmatic engineer should treat the following as plausible contents rather than confirmed facts:

• Performance and scheduling optimizations — improved graph partitioning, operator placement onto the Hexagon NPU (HTP), and enhancements to the QNN context binary caching that reduce model load time and first‑inference latency.
• Stability and operator coverage — fixes to operator mapping that previously led to CPU fallback or session creation failures for particular model topologies.
• Compatibility alignment with Qualcomm runtime versions — small ABI or binding adjustments so ONNX Runtime QNN EP plays nicely with QAIRT/QNN versions shipped on devices or used by Qualcomm AI Hub. Qualcomm’s release notes show frequent QAIRT upgrades and contextual changes.
• Input sanitization and robustness hardening — improvements to reduce parsing errors or malformed input corner cases (plausible, but not verifiable without vendor disclosure).

Important caveat: none of the above can be accepted as definitive until verified against vendor release notes, telemetry, or direct testing on the device families in question. The KB’s “includes improvements” statement is intentionally non‑specific.

---

Practical impact for users and administrators​

For end users (consumers)​

Most users on qualifying Copilot+ Qualcomm devices will see incremental, subtle improvements rather than headline features. Expected practical outcomes include:

• Slightly snappier responses for local Copilot flows that use on‑device SLMs or small LLMs.
• Improved image editing responsiveness and potentially cleaner segmentation/mask edges in Photos/Studio Effects.
• Reduced battery or CPU usage when NPU offload succeeds more often.

These improvements are device‑ and firmware‑dependent; on older Snapdragon variants or OEM images without matching drivers, changes may be negligible.

For IT administrators and enterprise rollouts​

Treat KB5067994 as a component‑level OS change that interacts with vendor drivers and firmware. Recommended operational posture:

• Pilot the update on a representative fleet of Copilot+ Qualcomm devices (7–14 days recommended).
• Capture pre‑update baselines for: time‑to‑first‑token, model latency for common SLM/LLM flows, NPU/CPU utilization and battery drain under representative workloads.
• Ensure OEM and Qualcomm drivers/firmware are current; many regressions stem from driver/runtime mismatches.
• If regressions occur, collect Event Viewer logs, reliability/WER buckets, and component Update history entries and escalate to OEM or Microsoft with diagnostic packages.

---

How to verify KB5067994 on a device​

• Open Settings → Windows Update → Update history. After the component installs, you should see an entry labeled something like: “2025‑09 QNN Execution Provider version 1.8.13.0 (KB5067994).”

For deeper runtime verification:

• Developers can interrogate ONNX Runtime programmatically with get_providers and review session logs to confirm the QNN EP is registered and used for a model run. Enable ONNX Runtime debug logs to surface operator mapping decisions and fallback behavior.
• Use targeted test models (for example, quantized CV models or representative SLM prompts) and measure time‑to‑first‑token and throughput with and without the QNN EP active. Record thermal and power telemetry to detect unintended regressions.

---

Developer guidance: model and runtime considerations​

• Quantization: QNN EP typically favors quantized models (QDQ, w8a8, w8a16 patterns). Validate your quantization pipeline and operator coverage on representative target devices. The ONNX Runtime docs and Qualcomm tools (QAIRT) provide quantization and context binary guidance.
• Context binary cache (ep.context_enable): enabling context binary dumping can materially reduce session creation time by caching compiled QNN graphs. Use with caution on devices with constrained storage and validate cache behavior across app updates.
• Fallbacks: test disable_cpu_ep_fallback in controlled tests if you need to verify full NPU mapping; otherwise, CPU fallback will mask mapping failures.
• Prebuilt packages: ONNX Runtime prebuilt packages for Windows may include QNN dependencies (post certain ORT versions), simplifying deployment for many apps that consume the system EP. Confirm the runtime version if you depend on a specific feature or bugfix.

---

Validation checklist for IT teams (recommended, step‑by‑step)​

• Verify prerequisites: confirm devices are Copilot+ Qualcomm hardware and run Windows 11 24H2 with the latest cumulative update.
• Inventory: select devices across OEMs and thermal designs representative of production.
• Baseline capture: measure model latency, tokens/sec for SLM flows, NPU/CPU utilization, and battery drain during representative tasks.
• Pilot install: allow Windows Update to apply KB5067994 in a controlled ring or deploy via Microsoft Update Catalog / management tooling. Monitor Update history to confirm package presence.
• Post‑install verification: re‑run baseline tests, validate Photos/Studio Effects/Windows Hello functionality, and monitor Event Viewer for NPU/driver errors or crashes.
• Rollout staging: if pilot is successful, expand rollout in waves while maintaining telemetry collection. If regressions occur, revert using system restore or pre‑update images and escalate with collected logs.

---

Strengths, risks and editorial assessment​

Strengths​

• Targeted per‑silicon tuning allows Microsoft and Qualcomm to iterate on NPU‑specific operator placement and runtime stability faster than monolithic OS updates. This agility benefits on‑device AI features that require tight latency budgets.
• Automatic Windows Update delivery simplifies consumer distribution and ensures many qualifying devices stay current without manual intervention.
• System‑wide benefit for apps using ONNX Runtime: improvements to the system QNN EP can help multiple apps without requiring each app to bundle vendor runtimes.

Risks and operational caveats​

• Opaque changelog and lack of CVE detail: Microsoft’s KB text offers no granular disclosures; security and compliance teams must treat the KB as operational change with unknown micro‑level impacts until further vendor notes are provided. Flag any unverified fix/performance claims.
• Driver/firmware mismatches: many regressions arise from vendor runtime and OEM driver mismatches rather than the EP package itself. Always validate OEM driver alignment before mass deployment.
• Rollback complexity: component updates that touch hardware acceleration stacks can be harder to rollback cleanly; maintain system images and documented recovery paths.

---

When to escalate to OEM or Microsoft support​

If post‑update regressions manifest as visual artifacts in image pipelines, camera capture errors, frequent crashes in ONNX Runtime sessions, or NPU driver faults, collect:

• Update history entry (component name and version).
• OEM driver and firmware versions.
• Windows Event logs and crash dumps.
• Repro steps, sample inputs (images, model files), and timings.

Escalate to OEM support if evidence points to camera/ISP or GPU driver interactions; escalate to Microsoft when the issue appears tied directly to the QNN EP package or ONNX Runtime behaviors that can be shown to differ pre/post update. Include the full diagnostic package to expedite triage.

---

How this fits into the larger on‑device AI strategy​

Microsoft’s modular AI component approach — shipping vendor EPs and model runtimes independently — reflects the practical realities of heterogenous NPUs and the need for frequent iteration on operator mappings, quantization heuristics, and runtime scheduling. It reduces friction for developers and helps deliver device‑optimized experiences faster, but it places the onus on IT and engineering teams to validate behavior across a fragmented hardware landscape. The QNN EP 1.8.13.0 update is one of many iterative steps in that journey.

---

Conclusion — recommendations and final takeaways​

KB5067994’s QNN Execution Provider update (1.8.13.0) is a purposeful, narrowly scoped component change intended to improve Qualcomm‑based on‑device AI on Copilot+ Windows 11 (24H2) machines. Allow the update on consumer devices and benefit from automatic Windows Update delivery, but for managed fleets follow a disciplined rollout: verify prerequisites, pilot on representative hardware, capture pre‑ and post‑update telemetry, keep OEM drivers up to date, and maintain rollback images.
Treat the KB’s high‑level wording — “includes improvements” — as a signal to validate empirically rather than accept unspecified performance or security claims. Use ONNX Runtime session logs, Qualcomm AI Hub/QAIRT tooling, and controlled benchmarks to confirm operator coverage, latency, and power/thermal behavior on your target devices. When in doubt, escalate with full logs to OEMs or Microsoft; the public KB is intentionally brief and may not contain the forensic details procurement or security teams sometimes require.

---

Appendix: Quick reference commands and checks

• Check Update history: Settings → Windows Update → Update history and look for “QNN Execution Provider version 1.8.13.0 (KB5067994).”
• ONNX Runtime: enable detailed session logging to observe provider registration and operator fallback behavior (see ONNX Runtime QNN EP docs).
• Qualcomm tooling: use Qualcomm AI Hub/QAIRT for model compilation and to inspect context binary generation behavior on target devices.

This update is small in scope but strategically meaningful: it tightens the runtime layer that underpins many on‑device Copilot experiences and underscores the evolving operational responsibilities for admins and developers in a world of modular, silicon‑specific AI components.

---

Source: Microsoft Support KB5067994: Qualcomm QNN Execution Provider Update (1.8.13.0) - Microsoft Support