KB5066125 Phi Silica Update: On-Device AI v1.2508.906.0 for Qualcomm Copilot+

Microsoft has pushed another incremental but important update for on‑device AI: KB5066125 upgrades the Phi Silica AI component to version 1.2508.906.0 for Qualcomm‑powered Copilot+ PCs, delivered automatically through Windows Update to qualifying Windows 11 (24H2) devices. (support.microsoft.com)

Background / Overview​

Phi Silica is Microsoft’s small language model (SLM) engineered to run locally on Copilot+ PCs by offloading inference to the device Neural Processing Unit (NPU). Its design goals — aggressive weight quantization, low idle memory, fast time‑to‑first‑token, and NPU‑first operator placement — make it the on‑device backbone for a range of Copilot experiences such as quick summarization, rewrite features and early multimodal image descriptions. These design points are explained in Microsoft’s technical posts and developer documentation, and they remain the reference for interpreting this KB. (blogs.windows.com, learn.microsoft.com)
The KB entry itself is brief: it confirms the version bump to 1.2508.906.0, states the update “includes improvements to the Phi Silica AI component for Windows 11, version 24H2,” and lists the distribution and prerequisite details. The release replaces a prior Qualcomm‑targeted component release. Administrative and end‑user guidance in the KB emphasizes automatic deployment via Windows Update and the requirement that the latest cumulative update for Windows 11, version 24H2 must already be installed. (support.microsoft.com)

---

What KB5066125 actually says​

The public facts (concise)​

• The update targets Copilot+ PCs running Windows 11, version 24H2 and is scoped to devices using Qualcomm NPUs. (support.microsoft.com)
• It installs automatically via Windows Update; the device must have the latest cumulative update for 24H2 before this component will apply. (support.microsoft.com)
• The update replaces the previously released Qualcomm Phi Silica package referenced by Microsoft. (support.microsoft.com)

What Microsoft does not publish​

Microsoft’s KB does not include a line‑by‑line engineering changelog: there are no public notes enumerating exact model weight changes, operator adjustments, quantization tweaks, or per‑operator runtime fixes. That omission is intentional for many component updates; it means administrators and engineers must infer impact from telemetry, OEM driver notes, and post‑install testing. Expect the KB’s terse wording: “includes improvements” rather than specific technical diffs. (support.microsoft.com)

---

Technical context: why platform‑specific Phi Silica updates matter​

Phi Silica’s design and published performance targets​

Microsoft has described Phi Silica as a Transformer‑based SLM tuned for NPU execution with concrete design targets that matter during rollouts:

• 4‑bit weight quantization for compact size and higher throughput.
• Time‑to‑first‑token target around 230 ms for short prompts.
• Sustained throughput targets on the order of up to ~20 tokens/sec under ideal conditions.
• Context window initially around 2k tokens, with extensions planned. (blogs.windows.com, learn.microsoft.com)

These figures are published by Microsoft as design targets and lab measurements and are useful planning guides for developers and IT teams; they are not guarantees for all OEM hardware or real‑world workloads. Device thermal profiles, firmware, and NPU generation all materially affect real performance. (blogs.windows.com)

Why updates are per‑silicon​

NPUs are heterogeneous across vendors and generations, and the inference path combines the model runtime, the NPU driver stack, and Windows AI runtime scheduling. Small changes in operator placement, quantization rounding, or memory management can cause differences in latency, throughput, and stability. That is why Microsoft ships separate Phi Silica component builds by silicon family (Qualcomm, Intel, AMD) rather than a single one‑size‑fits‑all package. Component updates like KB5066125 typically address per‑silicon operator scheduling, runtime fixes, quantization edge cases, and multimodal projector calibration. (learn.microsoft.com, support.microsoft.com)

---

User and IT impact — what to expect after installation​

For end users​

• Faster and smoother local Copilot responses: Users should notice snappier replies for short, on‑device tasks (rewrite/summarize, Click to Do UI flows) where Phi Silica runs locally. Improvements are usually subtle — incremental latency and stability gains rather than broad new features. (blogs.windows.com, support.microsoft.com)
• Improved offline/privacy behavior: On‑device inference reduces the amount of data sent to cloud LLMs for routine Copilot interactions, which benefits privacy‑sensitive workflows. Note that cloud fallbacks remain for heavy multimodal generation and complex tasks. (blogs.windows.com)

For IT administrators​

• Sequencing requirement: Confirm the target devices have the latest Windows 11 24H2 cumulative update before the Phi Silica component will deploy; otherwise Windows Update will not apply the component. (support.microsoft.com)
• Staged rollouts recommended: Because the component touches runtime and model paths that interact with device drivers and firmware, pilot the update on a representative set of Qualcomm devices before broad deployment. Monitor event logs and telemetry. (support.microsoft.com)
• Rollback complexity: Component updates that change runtime behavior can be difficult to remove safely. Organizations should rely on imaging and tested rollback plans (system restore points, pre‑update images) rather than ad‑hoc package removal.

---

Developer and OEM implications​

Developers​

The Windows App SDK exposes Phi Silica APIs (experimental channel) enabling developers to call the local model directly from apps, perform text transformations, or use built‑in Text Intelligence Skills (summarize, rewrite, text→table). Developers must test apps against the new Phi Silica binaries on Qualcomm devices to validate latency, tokenization behavior, and any multi‑modal changes introduced by component revisions. If an app depends on specific latency or memory profiles, retest and adjust batching/timeouts accordingly. (learn.microsoft.com)

OEMs and driver vendors​

Because Phi Silica execution relies on NPU drivers and firmware, OEMs must verify their Qualcomm driver stacks work smoothly with the updated OS component. Changes in quantization ranges, projector normalization for vision adapters, or memory management can uncover firmware edge cases; coordination between Microsoft, Qualcomm, and OEM engineers is often necessary. Historical rollout patterns show isolated device‑specific regressions tied to driver mismatches — these are uncommon but significant when they occur.

---

Multimodal capabilities and accessibility — practical notes​

Microsoft’s applied sciences team has deliberately extended Phi Silica with vision adapters (a vision encoder plus a relatively small multimodal projector) so image understanding can be supported without shipping a separate large vision SLM on device. This connector approach reuses existing encoders (e.g., Florence) and a small 80‑million‑parameter projector to translate image embeddings into Phi Silica’s embedding space, keeping the memory and disk footprint low. Multimodal image descriptions are already used for accessibility scenarios (alt text, detailed descriptions) where short captions run in ~4 seconds and longer descriptions in ~7 seconds on targeted hardware — numbers published by Microsoft and verified in engineering blog posts. These timings are useful benchmarks but subject to device variation. (blogs.windows.com)

---

Security and privacy analysis​

Privacy benefits​

• Reduced data egress: Local inference keeps user prompts and context on the device for many routine interactions, aligning with data‑residency and privacy goals for enterprises and individual users. (blogs.windows.com)

Security considerations and attack surface​

• Model in the trusted computing base: As model binaries and runtime become part of the device’s trusted base, organizations should treat them like firmware — ensure update channels are secure and that devices apply signed updates only. Microsoft signs these component updates and distributes them over Windows Update. (support.microsoft.com)
• Telemetry & diagnostics: Even when models run locally, diagnostic telemetry or cloud fallbacks may transmit metadata; validate Copilot and Windows privacy settings against organizational policy to avoid unexpected data flows.

Unverifiable claims flagged​

Public Microsoft materials publish design targets (e.g., 230 ms time‑to‑first‑token, up to ~20 tokens/sec). Those figures are lab results and company targets; real‑world throughput and latency should be validated on representative hardware. Any claim about exact internal changes in this KB (for instance, “we updated quantization from 4‑bit to 3.5‑bit”) is not verifiable from the KB itself and must be treated as speculative unless Microsoft or an OEM publishes detailed engineering notes. Flag: the KB does not disclose per‑operator or weight‑level changes, so any inference about exact internal changes remains unverified public conjecture. (blogs.windows.com, support.microsoft.com)

---

Troubleshooting, monitoring and rollback guidance​

Quick checks after deploying KB5066125​

• Confirm update presence: Settings → Windows Update → Update history should list “2025‑08 Phi Silica version 1.2508.906.0 for Qualcomm‑powered systems (KB5066125)”. (support.microsoft.com)
• Monitor Event Viewer for Copilot/AI runtime errors and kernel/gpu driver warnings. Track LiveKernelEvent or reliability monitor entries that could indicate driver interactions.
• Validate NPU and driver versions: ensure OEM Qualcomm drivers/firmware are at the versions Microsoft and the OEM recommend for Copilot+ certification. Driver mismatches are the most common cause of post‑update instability.

Diagnostic telemetry to collect​

• Time‑to‑first‑token and sustained tokens/sec for representative prompts.
• NPU utilization and CPU offload metrics under steady state.
• Battery and thermal telemetry across 10–30 minute sustained workloads.
• Any newly surfaced application crashes or freezes correlated with the update timestamp.

Rollback options and cautions​

• If severe regressions occur, prefer restoring from a pre‑update system image or using a verified system restore point.
• Manual package removal via DISM for LCUs can be complex and is not always reliable for component updates; Microsoft’s guidance and community experience advise caution.

---

How to validate the update in a lab (recommended checklist)​

• 1.) Identify representative Qualcomm Copilot+ devices across OOB thermal designs (thin laptop, convertible, and larger laptop).
• 2.) Capture pre‑update baselines: token latency, NPU/CPU utilization, battery drain, and reliability metrics for Copilot flows.
• 3.) Apply the prerequisite cumulative update, then allow the Phi Silica component to install via Windows Update.
• 4.) Rerun the same workload suite and compare deltas: time‑to‑first‑token, tokens/sec, and telemetry spikes.
• 5.) Monitor for new Event Viewer or reliability entries for 72 hours of typical usage.
• 6.) If regressions are observed, collect repro steps and escalate to OEM and Microsoft support with diagnostic logs.

---

Strengths, practical benefits and remaining risks​

Notable strengths​

• Performance and latency gains: Hardware‑tuned model updates continue to improve responsiveness for common Copilot tasks. (blogs.windows.com)
• Privacy and offline utility: Local model improvements reduce cloud dependency for many routine flows. (blogs.windows.com)
• Developer enablement: Windows App SDK integration lets apps leverage the updated local model without shipping their own model binaries. (learn.microsoft.com)

Principal risks and limitations​

• Opaque changelogs: The KB’s lack of granular detail complicates change management and incident triage for enterprises. (support.microsoft.com)
• Hardware fragmentation: User experience will vary by OEM device, NPU generation and firmware maturity; not all Qualcomm devices will achieve the same gains.
• Rollback and remediation complexity: Component updates can interact with drivers, making recovery nontrivial in some environments.

---

Editorial takeaway and practical recommendations​

KB5066125 is not a headline feature release; it’s an iterative tuning update in Microsoft’s broader on‑device AI rollout. For most users, the change will be invisible beyond modest snappiness and reliability improvements in local Copilot tasks. For IT professionals, the practical implications are clear: treat Phi Silica component updates like any OS‑level change that crosses into hardware acceleration territory — validate prerequisites, stage rollouts, collect targeted telemetry, and prepare tested rollback options.
Actionable checklist:

• Ensure target machines are confirmed Copilot+ and running Windows 11 24H2 with the latest cumulative update. (support.microsoft.com)
• Pilot KB5066125 on a small, representative device set and capture before/after performance and reliability baselines.
• Validate OEM Qualcomm drivers and firmware compatibility; apply OEM updates where recommended.
• If operating in a regulated environment, confirm telemetry and Copilot privacy settings meet organizational policy.

---

Conclusion​

KB5066125 (Phi Silica v1.2508.906.0) is another step in Microsoft’s platformization of on‑device AI: small, frequent component updates tune models to silicon, improve user‑facing latency and privacy, and reduce cloud dependency for routine Copilot experiences. The update is deliberately concise in public documentation — useful for broad adoption but leaving technical teams to validate outcomes through testing and telemetry. Organizations should adopt a measured rollout posture, prioritize driver/firmware compatibility checks, and treat these AI component releases as part of standard OS change management rather than optional gadgetry. (support.microsoft.com, blogs.windows.com, learn.microsoft.com)

---

Source: Microsoft Support KB5066125: Phi Silica AI component update (version 1.2508.906.0) for Qualcomm-powered systems - Microsoft Support

 

[ChatGPT]

Microsoft has pushed a platform-level Phi Silica update for Intel-powered Copilot+ PCs: KB5066126 upgrades the on-device Phi Silica AI component to version 1.2508.906.0, is delivered automatically through Windows Update, and requires the latest cumulative update for Windows 11, version 24H2 before it will install. (support.microsoft.com)

---

Background​

Phi Silica is Microsoft’s purpose-built, NPU‑tuned Small Language Model (SLM) designed to run locally on Copilot+ devices. It’s shipped as a managed Windows component so updates and model/runtime tuning can be deployed through Windows Update rather than requiring separate app-level model shipping. Microsoft positions Phi Silica to give a range of Copilot features — text rewrite, summarization, accessibility image descriptions and certain multimodal experiences — the ability to run on device NPUs with lower latency and reduced cloud exposure. (blogs.windows.com, learn.microsoft.com)
This new KB — KB5066126 — is the Intel-targeted wave of Microsoft’s ongoing component releases (parallel KBs have been issued for AMD and Qualcomm device families in prior waves). The KB entry confirms that the update replaces an earlier Intel build and will appear in Settings → Windows Update → Update history once installed. (support.microsoft.com)

---

What Microsoft’s KB actually says​

The KB’s public note is short and explicit about the essentials:

• The update applies to Copilot+ PCs running Windows 11, version 24H2 and the usual consumer and enterprise SKUs. (support.microsoft.com)
• The component update raises Phi Silica to version 1.2508.906.0 for Intel-powered systems and replaces the previously released Intel Phi Silica package. (support.microsoft.com)
• Distribution method: automatic via Windows Update. Prerequisite: the latest cumulative update for Windows 11, version 24H2 must be installed first. (support.microsoft.com)

Notably, the KB does not publish a line-by-line changelog. As with most on‑device model component updates, Microsoft summarizes the release as “improvements” rather than listing per-operator changes or detailed model diffs. That means administrators and developers must treat the KB as a versioning and distribution notice rather than as an engineering changelog. (support.microsoft.com)

---

Why this matters: Phi Silica’s technical context​

Phi Silica’s design goals and runtime characteristics​

Microsoft’s engineering posts and platform documentation make the model’s goals explicit: a compact, efficient SLM that favors NPU offload, 4‑bit weight quantization, low idle memory, fast time‑to‑first‑token, and practical context lengths to support common UI workflows rather than unconstrained long-form generation. Representative engineering numbers that Microsoft publishes for Phi Silica include a time‑to‑first‑token target on the order of ~230 ms for short prompts and throughput figures in the tens of tokens per second on supported NPUs; the current context length target is 2k tokens with near-term plans to scale to 4k in further updates. These are manufacturer-published lab figures and are useful planning targets — they are not universal guarantees across all Intel NPU implementations. (blogs.windows.com, learn.microsoft.com)

How Phi Silica fits into the Windows AI stack​

Phi Silica is delivered as part of the Windows AI components family and is consumable by system features and developers via the Windows App SDK (experimental channel for Phi Silica APIs). That means the same in‑box model is intended to power built‑in experiences (Copilot features, accessibility image descriptions, local rewrite/summarize flows) while being available to apps that call the Windows AI APIs. Microsoft’s Windows App SDK documentation and Windows Developer blog describe how these APIs and on‑device models let developers embed local AI without shipping large models themselves. (learn.microsoft.com, blogs.windows.com)

---

What KB5066126 likely changes — and what remains opaque​

Because the KB is terse, the community must infer the scope from Microsoft’s update patterns and the version bump:

• Expect performance and stability optimizations targeted to Intel NPUs and their driver stacks (operator placement, quantization tuning, memory footprints). Component releases historically tune the runtime and model connectors rather than adding major new user‑visible features.
• The update may include multimodal projector calibrations, small vision‑adapter tweaks, or tokenization adjustments that improve local image description quality or handling of specific workloads on Intel-based Copilot+ devices. Microsoft’s multimodal work reuses existing encoders and adds a small projector module rather than re-training the entire base model, so component updates often carry connector‑level tuning. (blogs.windows.com)

Unverifiable or omitted details:

• The KB does not disclose weight-level diffs, exact quantization parameter changes, model parameter counts, or per‑scenario telemetry. Those internal engineering details are not published in KB text and are therefore not independently verifiable from the KB alone. Administrators who need that level of detail should coordinate with Microsoft or OEM engineering channels and treat the KB as a public status update rather than a full technical report.

---

Deployment and enterprise operational considerations​

Prerequisites and delivery​

• Confirm devices are Copilot+ certified and running Windows 11, version 24H2. Only Copilot+ PCs will be offered the Phi Silica component. (support.microsoft.com)
• Install the latest cumulative update (LCU) for Windows 11 24H2 — the Phi Silica component requires this checkpoint before it will apply. (support.microsoft.com)
• Windows Update will download and install the component automatically on eligible devices; IT admins can control rollout using Windows Update for Business (WUfB), WSUS, or SCCM/Intune workflows. The Microsoft Update Catalog may publish CAB/MSU entries for offline staging, but not every component appears immediately in the Catalog.

Recommended staging checklist for IT​

• Pilot ring: apply KB5066126 to a small, representative group of Intel Copilot+ devices first, including different OEMs and firmware versions.
• Baseline telemetry: capture pre‑update baselines for Copilot latency, NPU utilization, and battery behavior under representative workloads.
• Driver validation: ensure the latest NPU/GPU drivers and firmware from OEMs are installed — driver mismatches are the most common cause of post-update anomalies.
• Logging and monitoring: enable diagnostic logs for AI components and monitor Event Viewer entries for AI runtime, NPU driver, or GPU errors.
• Rollback plan: have validated image-level recovery and driver rollback procedures. Component-only updates are not always trivially removable via the GUI; be prepared to restore system images if a critical regression occurs.

How to verify the update​

After Windows Update installs the component, check:

• Settings → Windows Update → Update history. The entry should read: “2025‑08 Phi Silica version 1.2508.906.0 for Intel‑powered systems (KB5066126).” (support.microsoft.com)

For large fleets, use Windows Update for Business reporting or Intune/Azure AD reporting to confirm installation status across rings.

---

Performance, privacy, and user impact​

Performance and battery profile​

Phi Silica’s engineering choices (4‑bit quantization, memory‑mapped embeddings, and operator placement optimized for NPUs) are explicitly intended to reduce runtime memory and power draw, enabling sustained interactive usage on thin devices. When inference runs primarily on the NPU, CPU and GPU remain available for other workloads, which can lead to improved thermals and battery life for common assistant tasks. However, the magnitude of performance gains depends heavily on the NPU silicon, drivers, OEM thermal profiles, and system memory. Microsoft’s throughput/time‑to‑first‑token numbers are lab‑measured benchmarks and should be validated against your hardware before being used as contractual claims. (blogs.windows.com, learn.microsoft.com)

Privacy and offline behavior​

One of Phi Silica’s primary advantages is local processing: for many Copilot interactions, user text and local content can be processed on device, minimizing data sent to cloud LLMs and improving privacy posture. That said, some Copilot features or enterprise configurations may still fall back to cloud models for heavier workloads or where cloud-only features are required. Administrators should review Copilot and Windows privacy settings and document which flows may route off‑device. Microsoft also includes built‑in content moderation and system-level Responsible AI tooling to control outputs and sensitivity — but app-level usage still requires developer and admin controls to ensure compliance. (blogs.windows.com, learn.microsoft.com)

---

Risks, limitations, and things to watch​

• Hardware fragmentation: Microsoft ships platform‑specific Phi Silica builds (Intel, AMD, Qualcomm). Feature parity and performance vary across NPUs. Expect different user experiences on different Copilot+ hardware classes.
• Opaque changelogs: the KB’s “improvements” language leaves important internal changes unspecified. Where organizations need reproducible behavior for compliance or auditability, that opacity is a real gap. Flag such updates for deeper engineering review in pilot windows.
• Potential regressions: component updates interact with drivers and firmware; historically, some AI component updates have triggered device‑specific issues requiring driver patches or rollbacks. Staging is essential.
• Rollback complexity: removing a component-level model update may not be a single-step GUI action in all scenarios; administrators should validate rollback procedures.
• Security and model integrity: shipping model binaries as part of the OS increases the trusted computing base surface. While Microsoft signs and distributes these components via Windows Update, organizations should ensure update channels (WSUS, WUfB) and code signing verification are enforced. Any tampering of model files or the runtime would be an attack vector.

Caveat on performance claims: independent third‑party benchmarks across the wide variety of Intel NPUs are currently limited compared with Microsoft’s lab numbers; treat published token‑rate and latency figures as vendor-published measurements and validate on your devices.

---

Developer implications​

• The Windows App SDK exposes Phi Silica APIs (experimental channel) so developers can call the in‑box model for text generation, summarization, and multimodal tasks. Apps using these experimental APIs should be tested against the updated Phi Silica binaries on target hardware, especially where timeouts, batching, and memory use are sensitive. (learn.microsoft.com)
• Microsoft announced LoRA support and other developer tooling to enable lightweight fine‑tuning scenarios for Phi Silica; such functionality is useful but experimental and carries governance implications for enterprises that plan to tune models on corporate data. Validate the lifecycle and rollback of fine‑tuned adapters before integrating them into production apps. (blogs.windows.com)
• Content moderation: Microsoft exposes content safety controls and system-level moderation models for Phi Silica. Developers must understand these controls and configure them appropriately for regulated or public‑facing applications. (blogs.windows.com)

---

Practical checklist — what to do next​

For home users:

• Ensure Windows Update is enabled and the machine is running Windows 11, version 24H2 with the latest cumulative update.
• Open Settings → Windows Update → Update history after patching and confirm: “Phi Silica version 1.2508.906.0 for Intel‑powered systems (KB5066126).” (support.microsoft.com)
• If you notice regressions in Copilot behavior after the update, capture repro steps and open feedback through the Feedback Hub.

For IT admins:

• Stage KB5066126 in a pilot ring that mirrors production hardware (varied Intel OEMs, drivers, firmware).
• Confirm the prerequisite LCU is applied before approving the Phi Silica component in WUfB/WSUS. (support.microsoft.com)
• Validate NPU/GPU drivers with OEMs and have rollback images/drivers ready.
• Monitor key metrics: Copilot latency, NPU utilization, battery draw, Event Viewer warnings. Log issues to OEM support and Microsoft Support if needed.

For developers:

• Test your apps against the updated Phi Silica runtime and monitor for subtle API behavior changes (timeouts, batch sizing, streaming semantics) before releasing updates that depend on local model behavior. (learn.microsoft.com)

---

Broader editorial analysis: strengths and long-term implications​

Microsoft’s approach — shipping small, efficient models as OS components that are NPU‑optimized and centrally managed — is strategically sound for making on‑device AI broadly available. The pattern of delivering incremental component updates allows rapid iteration on model quality, performance, and connectors without waiting for large OS feature updates. The advantages are clear:

• Lower latency and improved UX for many Copilot flows because local inference eliminates network round trips. (blogs.windows.com)
• Better privacy posture for many use cases by keeping more user context local. (blogs.windows.com)
• Developer convenience by exposing a consistent in‑box model via the Windows App SDK rather than forcing each app to ship and manage large models. (learn.microsoft.com)

However, the strategy also surfaces systemic risks that organizations must reckon with:

• Feature fragmentation across OEMs and NPU vendors will likely persist as silicon and drivers mature at different rates. This leads to inconsistent experiences and increased management overhead for IT.
• Opacity in update details puts pressure on engineering teams and compliance officers who require auditable change logs and reproducible behavior for regulated workflows.
• Operational complexity grows when the OS also becomes a vehicle for model delivery — update cadence, rollback, and forensic analysis of model behavior all become part of standard patch management disciplines.

Independent reporting and early reviews have underscored both the promise and friction of this approach: the gains in responsiveness and privacy are compelling, but driver/firmware readiness and the diversity of NPU capabilities remain gating factors for achieving uniform parity across the Windows ecosystem.

---

Conclusion​

KB5066126 is a routine but strategically significant step in Microsoft’s on‑device AI roadmap: it updates Phi Silica to version 1.2508.906.0 for Intel‑powered Copilot+ PCs and will be distributed automatically via Windows Update after the device has the latest Windows 11 24H2 cumulative update. The public KB is intentionally concise — it confirms the version, scope, distribution, and prerequisite but does not list granular model or runtime changes. Administrators, developers, and power users should treat the release as a versioning and delivery announcement and validate real‑world behavior on representative hardware before broad deployment. (support.microsoft.com, blogs.windows.com)
Practical next steps: stage the update in pilot rings, verify drivers and firmware from OEMs, capture performance and stability baselines, and confirm that privacy and content‑moderation controls meet organizational policy. The update continues Microsoft’s push to make local AI practical and private for endpoint scenarios — a clear architectural direction that brings both opportunity and new operational responsibilities.

---

Source: Microsoft Support KB5066126: Phi Silica AI component update (version 1.2508.906.0) for Intel-powered systems - Microsoft Support