fonnx 0.2.2

ONNX model inference for Dart and Flutter on every platform: ONNX Runtime through FFI on native, Web Workers on the web. Embeddings, Whisper, Kokoro TTS, and more.

0.2.2

• Requires code_assets 2.2, hooks 2.2 and kiln 0.2.2. kiln 0.2.0 left a .lock file next to each model that it downloaded, and the model catalog of 0.2.1 was not formatted for pub.dev. With code_assets 1.x, the package could not be installed with current objective_c and shared_preferences_foundation.
• The iOS 15 and macOS 14 floors for prebuilt selection compare the complete deployment version.
• Shorter package description, and a usage example for pub.dev.

0.2.1

• The prebuilt native libraries are used again: the 0.2.0 manifest was made before its version bump, so its source key did not match and every build compiled from source.

0.2.0

• The native libraries are rebuilt from the same sources, in release native-58359725c0fd7164. The package no longer has a Codemagic configuration, which changed its source key.

0.1.2

• ONNX Runtime 1.31.1 (from 1.27.0) on every platform, including the iOS build and ONNX Runtime Web. The bindings use C API 31. On an M4 Max, inference is the same speed or faster: NeMo CTC 24%, Whisper tiny 15%, MiniLM 9% and the keyword spotter 6% faster.
• The keyword spotter no longer turns off KleidiAI: 1.31.1 fixes the convolution bug (microsoft/onnxruntime#28571) that made it emit blanks.
• Building from source on Windows no longer fails when ProgramFiles(x86) is not set.

0.1.1

• The prebuilt native libraries are used again: in 0.1.0, files left out of the package still counted in its source key, so every build compiled from source.

0.1.0

• First release on pub.dev, MIT licensed.

0.0.1

• Use kiln, the new name of native_prebuilt, from pub.dev. nativePrebuiltRuntimeFiles is now prebuiltRuntimeFiles, and the web files load from kiln/fonnx/ (FonnxWeb.runtimeUri). The release workflow is kiln_release.yml, from Telosnex/telosnex_ci_tools.
• Remove ONNX Runtime Extensions. No model needs a custom operator, so the native assets are ONNX Runtime and the session finalizer, and the release builds no Extensions. createOrtSession no longer takes includeOnnxExtensionsOps, and fonnxOrtExtensionsDylibPathOverride is gone. On iOS, the Run Script phase aligns only onnxruntime.framework.
• Add SenseVoice: speech-to-text with Alibaba's SenseVoice Small (239 MB, int8 sherpa-onnx export): Mandarin, Cantonese, English, Japanese and Korean, with the language detected or set. Native and web. Kaldi filter bank features, greedy CTC and token decoding are in shared Dart. On FLEURS: 5.5% CER Mandarin, 6.4% Japanese, 7.7% Korean, 7.0% WER English (Whisper Small: 9.8, 12.3, 8.9, 6.2). The NeMo log-mel features and SenseVoice share lib/src/power_spectrum.dart.
• The model files moved from Git and the GitHub runtime release to Hugging Face telosnex/fonnx. nativePrebuiltRuntimeFiles pins each file by Hugging Face commit and SHA-256, and now also has the four NemoAsr files. tool/fetch_hf_models.dart downloads the models for the tests and the example app. The old GitHub release stays for earlier fonnx commits.
• Add NemoAsr: speech-to-text with NVIDIA's multilingual FastConformer CTC (10 languages, 132 MB) or Parakeet TDT 0.6B v3 (25 European languages, 670 MB). Both are the int8 exports of sherpa-onnx, downloaded from Hugging Face. Native and web. Log-mel features, greedy CTC and TDT decoding, and token decoding are in shared Dart. On FLEURS, the CTC model is about as accurate as Whisper small and 3x faster than Whisper tiny; v3 is more accurate than Whisper small outside English.
• Whisper runs the split fp32 export of onnx-community/whisper-tiny (encoder, and merged decoder with a key/value cache) on stock ORT, native and web. Log-mel features, a 2-beam search with Whisper's token suppression, and token decoding are in shared Dart. ORT Extensions are no longer needed for Whisper. Whisper.load takes a WhisperBundle; Whisper.removeTimestamps is removed because timestamp tokens are suppressed.
• Use native_prebuilt for the native hook, shared verified cache, source-key selection, and build modes. Release CI builds the session finalizer through the hook for all ten targets. Consumer builds no longer compile the finalizer. Upstream ONNX Runtime URLs stay unchanged.
• Publish the 16 example ONNX models as verified runtime downloads. The runtime_models.dart API exposes their pinned catalog and download helper.
• Build Linux Extensions on Ubuntu 22.04 (glibc 2.35) instead of Ubuntu 24.04.
• Make native_artifacts/manifest.json the canonical source for all ten native target pairs, ORT/Extensions/Web source pins, Web assets, and model fixtures.
• Align native and Web on ONNX Runtime 1.27.0, remove mixed 1.17/1.19 workers, vendor the Web runtime locally, and remove executable CDN dependencies.
• Consume every owned OrtStatus exactly once and clean partial session state on all model/session construction failures.
• Add an explicit package finalizer ABI and an 8,192-case ownership corpus under AddressSanitizer and UndefinedBehaviorSanitizer.
• Snapshot mutable audio, token, byte, state, and keyword inputs at API entry; validate hostile keyword/path/search configuration in release builds.
• Add 21 deterministic frontend recipes, exact model/runtime verification, real local Web identity and Magika Worker inference, a fatal-safe versioned Worker RPC, and one authoritative serialized gate.
• Runtime-test exact archives on macOS, iOS Simulator, Android API 36, Linux arm64/x64, and Windows/Wine; record the existing Linux Extensions floor as glibc 2.38/GLIBCXX_3.4.32 instead of implying glibc 2.31 compatibility.