Intel has published v1.35.0-rc1 of its Linux NPU driver, delivering the first public preview of support for Wildcat Lake, the chipmaker's upcoming next-generation AI processor. The 90-file update also bumps the Level Zero dependency to v1.28.2, modernizes the NPU compiler API, and replaces external build submodules with local copies for improved build reproducibility. Developers will find a new dynamic shape optimization flag for generative workloads alongside thread-aware tracing and a freshly gutted stats subsystem. The release candidate is now available on GitHub for those looking to validate Wildcat Lake silicon ahead of hardware.

Intel Releases Linux NPU Driver v1.35.0-rc1, Points to Wildcat Lake in New Firmware

The release candidate updates the compiler, bumps Level Zero, and adds the first firmware binaries for Intel's next-gen AI processor.

Intel published v1.35.0-rc1 of its Linux NPU driver on July 22, and it's the first public preview since v1.33.0 landed back in early June. If you're tracking Intel's upcoming AI silicon, the headline isn't the version number. It's the 50xx firmware family.

The release touches 90 files across five major subsystems. Built on four commits, it's not just a calendar-based bump. The team at Intel moved the bundled NPU compiler to npu_ud_2026_28_rc1, bumped Level Zero to v1.28.2, and refreshed the firmware binaries for VPU 37xx, 40xx, and 50xx platforms. On top of that, the build system got a substantial overhaul aimed at reproducibility.

Wildcat Lake is clearly on the calibration table

The new 50xx firmware covers Panther Lake and Wildcat Lake. Wildcat Lake is Intel's next-generation AI PC processor, slotted to succeed Panther Lake in the Core Ultra lineup. That firmware update, paired with a telemetry bug fix in intel-npu-smi and new scheduling barrier types in the VPU headers, tells you exactly where things stand. WW28 sits right in Intel's typical pre-launch driver certification window, which usually precedes a late-2026 or early-2027 consumer rollout. Keep in mind that firmware binaries for the 50xx family actually got a little smaller in this build, dropping 6,576 bytes from the 1.33.0 size. That's not a regression. It's just what happens when you strip out PMT workarounds for early silicon and tighten the telemetry scaling.

The compiler API is getting its own spring cleaning. The deprecated vclAllocatedExecutableCreate3() is being replaced by vclAllocatedExecutableCreate4(), which returns an executable handle directly without forcing callers to juggle compatibility buffers. A new vclExecutableGetCompatibilityString() query rounds it out, letting applications pull version strings tied to compiled models. It's the kind of polish that matters when you're distributing NPUs across multiple hardware generations.

On the dependency side, Level Zero jumped to v1.28.2, aligning with the v1.15.31 oneAPI specification. OpenVINO, ONNX Runtime, and the GenAI subproject all moved to newer commits pinned to the WW28 cycle. The validation layer picked up a performance tracing component and Python sysman bindings, which should help anyone trying to profile NPU workloads programmatically. Magdalena Schulfer's release commit officially tags the build as the "unified 2026WW28" drop, which tracks with Intel's standard cadence for hardware readiness signaling.

Build and inference updates

Developers tracking the GitHub repo will notice the Perfetto submodule was yanked entirely. Its contents now live as a local copy under third_party/perfetto/sdk/. Three new CMake flags force the build to pull dependencies from submodules instead of relying on system packages. Ubuntu 26.04 PPAs are now supported alongside 24.04. If you've ever fought with broken CI pipelines or missing LFS objects, this is the fix you've been waiting for.

The user-mode driver layer got a handful of functional changes. VPUDriverApi::getBufferInfo() now returns a PAGE_SIZE-aligned buffer size directly. The inference execute path accepts a new optimizeForDynamicShapes flag, which should help generative AI workloads that lean on variable tensor dimensions. Disk cache initialization was simplified, and the tracing layer finally logs [tid:<thread-id>] in its Level Zero output. The old stats subsystem was gutted, roughly 143 lines of stats.cpp and stats.hpp are gone, and it's almost certainly replaced by Perfetto-based tracing now that the SDK is inlined.

A new 214-line graph profiling validation test was added to the UMD suite. It sits alongside existing tests for command lists, caching, external memory, tensor strides, and CID.

It's a rather meaty release candidate for what's essentially a driver stack. The Wildcat Lake groundwork is exactly what you'd expect this far ahead of launch, and the Level Zero alignment keeps the broader oneAPI ecosystem in sync. However, at the same time, the lack of WSL2 support in the open issues is starting to feel overdue for a Linux-first driver. Issue #56 has been sitting open for a while, and the UMq fast path is still marked as a work in progress.

The full diff is available on GitHub, and the release is tagged fd49947d. Build it yourself if you want to run Wildcat Lake validation ahead of hardware, or wait for distro packaging if you're on Intel silicon today.

Head here to the release tag and the full commit diff.