anira-backends

Prebuilt inference-engine binaries for anira. Each backend is built from source (or repackaged from an upstream prebuilt) and published as GitHub release archives that anira's CMake downloads at configure time.

Backends

Backend	C API	Upstream	lib name	License
TFLite	TfLite*	tensorflow/lite	tensorflowlite_c	Apache-2.0
LiteRT	LiteRt* ²	google-ai-edge/LiteRT	LiteRt	Apache-2.0
ONNXRuntime	—	microsoft/onnxruntime	onnxruntime	MIT
LibTorch	—	pytorch/pytorch	torch	BSD-3

TFLite and LiteRT are two C APIs for the same runtime (LiteRT is the rebranded TensorFlow Lite): TfLite* is the mature legacy API; LiteRt* is LiteRT's newer native API. Pick one.

This repo is licensed Apache-2.0; the published binaries follow their upstream licenses (above).

Support matrix

What ships per target — shared and/or static:

Target	TFLite	LiteRT ²	ONNXRuntime	LibTorch
macOS x86_64	shared · static	shared · static	shared · static	shared
macOS arm64	shared · static	shared · static	shared · static	shared
macOS universal	shared · static	shared · static	shared · static	shared
Linux x86_64	shared · static	shared · static	shared · static	shared
Linux aarch64	shared · static	shared · static	shared · static	shared
Windows x86_64	shared · static ¹	shared · static ¹	shared · static ¹	shared
Windows arm64	shared · static ¹	shared · static ¹ ³	shared · static ¹	shared
Android (arm64-v8a + x86_64)	shared · static	shared · static ⁴	shared · static	—
iOS (xcframework)	static	shared	static	—

macOS shared dylibs are Developer ID code-signed (Hardened Runtime, timestamped); the consuming app re-signs/notarizes on embed.

¹ Windows static also ships a Debug variant.

² LiteRT ships the native LiteRt* C API (libLiteRt) — distinct from the tflite engine's legacy TfLite* API. shared is repackaged from official prebuilts (litert/prebuilt/, pinned to a main SHA) plus from-source Bazel where none ship (macOS x86_64, Windows-arm64). static is built from source (no static lib ships upstream) and merged into one archive. Only iOS static isn't provided (iOS ships the prebuilt xcframework).

³ Windows-arm64 has no upstream prebuilt — both shared (libLiteRt.dll) and static are built from source with clang-cl + LLVM 20 (MSVC cl can't compile the deps' GCC/clang constructs), without XNNPACK (its pinned Bazel build has no arm64-Windows microkernels) — CPU via ruy/builtin.

⁴ Android static is built from source inside LiteRT's public ml-build container (which provides the cuda_redist/rules_ml_toolchain externals a bare runner lacks), then merged per-ABI.

— = not provided.

Releases

Backends are versioned independently but released together, keyed to the anira version: tag v2.0.3 builds every backend at its pinned engines/<backend>/VERSION and publishes all archives to a single release v2.0.3.

Sponsor