ExecuTorch

license

BSD-3-Clause(OSI)

source

public(the published repository is the engine)

core-gated

ungated(no enterprise path in the repository root and no paid build of the engine in the README)

The LICENSE body is the BSD-3-Clause text, read in full rather than taken from the API's label, and carries no appended condition. The repository is public and unarchived and builds the engine itself. Its root tree carries no enterprise, ee or commercial directory and its README describes no licence-gated build, so the core reads as ungated. Confidence is medium because that is a repository-and-README read rather than a pricing-page read.

• https://raw.githubusercontent.com/pytorch/executorch/main/LICENSE recorded 2026-08-31

  LICENSE body, read in full: the stock BSD 3-Clause text carrying eight copyright holders - Meta, Arm, Qualcomm, Apple, MediaTek, NXP, Samsung and Intel - which is why the API reports NOASSERTION. No appended condition.

• https://api.github.com/repos/pytorch/executorch recorded 2026-08-31

  Repo metadata for pytorch/executorch - archived false, 4,969 stars.

• https://raw.githubusercontent.com/pytorch/executorch/main/README.md recorded 2026-08-31

  README, read for a paid tier, an enterprise edition or a licence-gated build beside the published source, and for whether any component is stated to be private.

Banded on reported deployment rather than on stars, because stars measure the wrong thing for a runtime that ships inside other companies' applications. Meta's own engineering blog and the PyTorch documentation both state that ExecuTorch powers on-device inference across Instagram, WhatsApp, Messenger, Quest and Ray-Ban Meta smart glasses, for billions of people. The LICENSE corroborates the hardware side independently: its BSD-3-Clause header carries eight copyright holders - Meta, Arm, Qualcomm, Apple, MediaTek, NXP, Samsung and Intel - which is vendor participation in the runtime itself rather than a vendor list in marketing copy. `reported_traction` records a word and never a number, so no user figure is asserted here; 4,969 GitHub stars would have banded this at 2 and understated it by three levels.

• https://engineering.fb.com/2025/07/28/android/executorch-on-device-ml-meta-family-of-apps/ recorded 2026-08-31

  Meta's engineering blog on ExecuTorch in the Meta family of apps, naming Instagram, WhatsApp, Messenger and Quest and describing on-device inference reaching billions of people.

• https://docs.pytorch.org/executorch/main/success-stories.html recorded 2026-08-31

  The project's own success-stories page: "Billions of users", naming Instagram, WhatsApp, Messenger, Quest and Ray-Ban deployments.

• https://raw.githubusercontent.com/pytorch/executorch/main/README.md recorded 2026-08-31

  README: "It powers Meta's on-device AI across Instagram, WhatsApp, Quest 3, Ray-Ban Meta Smart Glasses" and "Powers billions of users at Meta with real-time on-device inference".

One band below the vllm anchor, level with onnx-runtime, which is banded there on the same instrument: execution-provider breadth across CPU, GPU, mobile and edge. An earlier reading put this two below on the view that an on-device runtime is narrow by nature, which had it backwards - the backend matrix here spans eight silicon vendors and four host platforms, which is wider than apple-core-ml-runtime's single-vendor coverage at the same band. It stays below the anchor because it executes exported models on device rather than serving them with batching, parallelism and a multi-tenant API surface.

• https://raw.githubusercontent.com/pytorch/executorch/main/README.md recorded 2026-08-31

  README, read for the engine's feature set: On-device execution of exported models through a ~50KB C++ runtime, with an export and lowering toolchain and backends across a dozen hardware vendors.

• https://raw.githubusercontent.com/pytorch/executorch/main/README.md recorded 2026-08-31

  README backend matrix and "12+ Hardware Backends - Open-source acceleration for Apple, Samsung, Qualcomm, ARM, MediaTek, Vulkan, and more", with the per-platform table naming XNNPACK, Vulkan, Qualcomm, MediaTek, Samsung Exynos, CoreML, OpenVINO, CUDA, Metal and MLX.