flutter_edge_ai

ONNX Runtime GenAI 0.17.0 for Android arm64, built without telemetry, for flutter_edge_ai_onnx. Its hook/build.dart fetches this zip at build time and checks its SHA-256.

Why not Microsoft's AAR. The libonnxruntime-genai.so in Microsoft's 0.17.0 Android AAR lists libmat.so (the 1DS telemetry SDK) in DT_NEEDED. That libmat.so is 4 KB-aligned. Google Play rejects it for apps that target API 35 or later, and a device with 16 KB pages cannot load it.

What this build is.

• Source: microsoft/onnxruntime-genai v0.17.0, commit 1d67be30e5b1ed0c264f6f41c598f85157e8288b.
• Built against ONNX Runtime 1.27.0: headers and libonnxruntime.so from Maven Central com.microsoft.onnxruntime:onnxruntime-android:1.27.0, SHA-256 checked.
• CMake: ENABLE_TELEMETRY=OFF, Release, arm64-v8a, Android API 27.
• Toolchain: Android NDK r29 (29.0.14206865, clang 21.0.0).
• Script: packages/flutter_edge_ai_onnx/native/build_genai_android.sh.

What was checked on this file.

• DT_NEEDED: liblog.so, libm.so, libdl.so, libc.so only.
• Every PT_LOAD segment has p_align 0x4000, and the hook's own check (requireAndroidPageAlign) accepts it.
• The library contains no OneCollector string and no 1DS endpoint.

Asset.

The build is not byte-reproducible across machines: the NDK version, the build directory (in __FILE__ strings) and the zip entry times all change the bytes. The SHA-256 above is the pin in hook/build.dart.

The -a suffix follows this repository's convention for a build of an upstream release that is not upstream's own binary.

The library is MIT-licensed (microsoft/onnxruntime-genai); ONNX Runtime, which it loads, is MIT-licensed (microsoft/onnxruntime).

Run on a device. Galaxy S25 Ultra (Android 16), through flutter_edge_ai_onnx with this library in the APK: Qwen2.5-0.5B-Instruct (ORT-GenAI) loads on the CPU, generates, stops mid-turn, answers a third turn in the same session and closes.

native-v0.18.0-g with the Windows archive changed alone.

Windows: both copies of libLiteRtWebGpuAccelerator.dll (lib-prefixed and not) are Google's prebuilt from LiteRT-LM 5f973f57 (sha256 2fb3583681ffb28d2fa8c08803445839e579a8e67417bde480415a4360e18ed0), which contains LiteRT #11120: on an Intel GPU the dp4a int8 fully-connected/convolution path is turned off. Gemma 4 12B on an Intel GPU under D3D12 returned no text for prompts under 512 tokens (LiteRT-LM #4083); it now answers.

Performance, Gemma 4 12B on an Intel GPU: prefill takes about 45% longer, because the fixed path runs in fp16/fp32 instead of int8. Measured on Panther Lake (Core Ultra X7 358H, Arc B390):

CPU and NPU do not load this library: Gemma 4 E2B on CPU/NPU, Intel's 12B on the NPU and EmbeddingGemma 2 (270m on CPU and _Intel_PTL on the NPU: same similarities to three decimals, same load times) measured unchanged. Other GPU vendors on Windows were not measured.

The accelerator imports nothing from libLiteRt.dll; its webgpu_dawn.dll imports resolve against the shipped Dawn, byte-identical at both upstream refs. The other 23 Windows files and the six other archives are byte-identical to native-v0.18.0-g.

macOS, Swift Package Manager: the two companion zips that macos/flutter_edge_ai_litertlm/Package.swift fetches are packaged for this tag from its macOS archive (the same bytes as -g's: same LC_UUID). Their CFBundleVersion is now numeric (0.18.8), the tag is in DevFlutteredgeNativeTag, and the zips no longer depend on the packager's timezone. checksums_spm_macos.txt lists them.

native-v0.18.0-f with the Windows archive changed alone: the Intel NPU stack moves from OpenVINO 2026.3.1 to 2026.4.1, and openvino_intel_npu_vm_runtime.dll joins it (#702). Same pins as -f: LiteRT-LM b2f686e2 (v0.18.0), LiteRT 26895c9f, Qualcomm dispatch against QAIRT 2.50.0.

What changed, per archive.

• Windows: LiteRtDispatch.dll, LiteRtLm.dll, StreamProxy.dll, the five OpenVINO DLLs and the four TBB DLLs come from CI run 39573062034 (commit 79649ded); openvino_intel_npu_vm_runtime.dll is new. The other twelve files are byte-identical to -f.
• Android, iOS, iOS simulator, macOS, Linux x86_64, Linux arm64: the -f archives, byte for byte (same sha256).

OpenVINO 2026.4.1. patches/litert_openvino_2026_4_1.patch is upstream LiteRT 991d15588a, applied on the pin until the pin contains it. The 2026.3 NPU plugin refuses Intel's Gemma 4 12B for Panther Lake: "Imported blob metadata version: 2.7 but the current version is: 2.6". openvino_intel_npu_vm_runtime.dll is loaded by the NPU plugin by name; Gemma 4 E2B and the 12B load without it, and I ship it because Intel's own kit does.

Verified with the bytes in this Windows archive. Probe app on the public API (Flutter 3.47.4 profile build), and litertlm_ffi_test with the sha256 of all 25 archive DLLs in the built app equal to the archive's:

native-v0.18.0-e plus four changes: the stream proxy takes its chunk accessors from Dart, Windows Intel NPU loads its model again, a stopped turn can be rewound in place, and the Android archive carries the Google Tensor NPU dispatch. Same pins as -e: LiteRT-LM b2f686e2 (v0.18.0), LiteRT 26895c9f, Qualcomm dispatch against QAIRT 2.50.0.

What changed, per archive. Every file not listed here is byte-identical to -e, Android's libLiteRtDispatch_Qualcomm.so and Linux arm64's Qualcomm dispatch and libcdsprpc.so shim included:

• All seven archives: libLiteRtLm and StreamProxy.
• Windows: LiteRtDispatch.dll, rebuilt from the pin.
• Android: libLiteRtDispatch_GoogleTensor.so, a new file.

stream_proxy_create2 (#668, #453). The Dart client resolves the three stream-chunk accessors on the libLiteRtLm it drives and passes them in, so the shim keeps no global state. stream_proxy_create is still exported, unchanged, for older Dart.

Windows Intel NPU (patch_c_api.sh §15, #679, LiteRT-LM #4023).

• Since LiteRT-LM v0.17.0, a bundle with a TF_LITE_AUX section (every AOT NPU bundle) was loaded through a C runtime file descriptor.
• Google's prebuilt libLiteRt.dll links its own C runtime, so it cannot read that descriptor; on v0.18.0 the process ended with 0xC0000409.
• Windows now loads those sections from memory, as v0.16.0 did.

Rewind a stopped turn (patch_c_api.sh §16, #682). New export: litert_lm_conversation_rewind_cancelled_turn. After a stopped generation, the conversation returns to its state before that turn, so earlier images and audio stay in context.

Google Tensor NPU dispatch (#670, #686).

• libLiteRtDispatch_GoogleTensor.so is built from LiteRT's litert/vendors/google_tensor/dispatch at the pin; patches/litert_google_tensor_dispatch_oss.patch drops an unused internal dependency.
• It loads the Pixel's own libedgetpu_litert.so.
• flutter_edge_ai_litertlm bundles it only for an app that sets google_tensor_npu: true.
• The directory it is built from is covered by Google's Tensor SDK terms (TERMS_OF_USE.md there) besides the sources' own headers.

Verified with the libraries in these archives. Provenance per run was checked by sha256 on Linux and Windows, LC_UUID on Apple, and ELF build ID in the APK on Android.

¹ One case, the same on both OpenCL GPUs: after a stopped audio turn the model said no city was mentioned, where the test expected the city on every GPU (calibrated on Metal, which passes it on the iPhone). The audio is there: in a probe on a Galaxy S24 (Adreno 750) the model, asked next to repeat the recording, wrote it word for word, and a stop at "OK." or no stop gave the city. The test now checks Android's GPU as it checks the CPU (#688).

NPU:

• Intel Lunar Lake 258V: NPU group 3 passed, 2 skipped (no Windows bundle with vision/audio towers); engine_create 0.8–1.1 s, ~61 chunks/s, greedy run1 == run2, no 0xC0000409.
• Qualcomm SM8750 (Galaxy S25 Ultra): NPU group and NPU vision/audio (towers on