Changed #

• The Apple Neural Engine (NeuRT) and Hexagon NPU (QHexRT) engines now ship from pinned, prebuilt archives published by the private engine repository rather than being compiled from source. Each archive is verified by SHA-256 and by a build receipt naming the plugin ABI version it was built against.
• The Hexagon engine's QAIRT/QNN runtime dependency is now pinned and fetched the same way — as a private, checksum-verified artifact from the engine repository, paired against the engine's own build identity so the two can never silently drift apart.
• No Flutter-facing API change.

Changed #

• The Apple Neural Engine (NeuRT) and Hexagon NPU (QHexRT) engines now ship from pinned, prebuilt archives published by the engine repository rather than being compiled from private source. Each archive is verified by SHA-256 and by a build receipt naming the plugin ABI version it was built against, so an engine built against a different ABI is refused at download time instead of relinking cleanly and corrupting dispatch at runtime.
• No Flutter-facing API change. The effect is on provenance: the Apple xcframework and the Android arm64-v8a payload this package depends on are now traceable to a published artifact and its checksum.

Changed #

• Version-train bump only; no Flutter-facing changes. This release folds the Electron binding's native build/package pipeline into the SDK's shared release workflow (release.yml) for the first time — no effect on Flutter.

Changed #

• Reissued the 0.20.23 runtime and model-catalog train from the corrected, immutable release manifest after repairing Linux WebGPU packaging.

Changed #

• Aligned the native runtime train on ONNX Runtime 1.28.0, Sherpa-ONNX 1.13.5, the maintained RunAnywhere llama.cpp fork, and the compatible MLX Swift line.
• Added current model catalog support and release-coherence validation across the SDK packages and example applications.

Changed #

• Version bump only. This release exists so the Electron packages can be republished with their Windows native staging fixed: 0.20.21 shipped the ONNX and Sherpa backends without the vendor runtimes they link, so those two engines could not load on Windows x64. The Dart/Flutter surface is unchanged from 0.20.19 and no 0.20.22 native artifacts are produced for it.

Changed #

• Version bump only. This release exists so the Electron packages can be republished carrying Windows x64 and Windows ARM64 natives; the Dart/Flutter surface is unchanged from 0.20.19 and no 0.20.21 native artifacts are published.

Changed #

• Version bump only. This release exists so the Electron packages can be republished with the staging telemetry URL baked in; the Dart/Flutter surface is unchanged from 0.20.19 and no 0.20.20 native artifacts are published.

• Commons: is_model_loaded now answers from the lifecycle store's READY state rather than matching on model_id alone, so a model whose load FAILED is no longer reported as loaded and no longer becomes undeletable.
• Commons: rac_model_registry_update_download_status allocates the replacement download path before mutating the registry, so a failed allocation preserves the existing path instead of losing it.
• Commons: streaming results report RAC_ERROR_OUT_OF_MEMORY on a failed string allocation instead of returning success with a null field.

• Commons: backend registration failures are isolated. An engine that cannot register is recorded as unavailable instead of aborting SDK initialization, so the remaining backends keep serving.
• Suite version bump. This release also rebuilds the Electron backend plugin natives; that part does not affect this package.

• Electron QHexRT backend and its commons platform-capability changes.
• Swift package is now consumable by version (from:) rather than only by git revision.

• Suite version bump.

• Suite version bump.

• Suite version bump.

• Suite version bump; Kotlin/QHexRT Maven release.

• Version-aligned release with benchmark aggregation, variance reporting, and updated model catalogs.

• CoreML diffusion enablement + macOS backend slice fixes (see SDK release notes).

Changed #

• Version-aligned 0.20.9 release across all RunAnywhere SDKs; dev-analytics config baked into the native core at build time.

Changed #

• Adopted the canonical 0.20 API contracts and release-native packaging.

Removed #

• Removed deprecated compatibility configuration APIs.

Changed #

• Added request-scoped native cancellation for voice-agent turn streams and aligned generated validation with the canonical IDL.

Removed (BREAKING) #

• VAD callback API: Removed RunAnywhereVAD.activityStream, setSpeechActivityCallback, setAudioBufferCallback, and setStatisticsCallback. Subscribe to RunAnywhere.vad.streamVAD(audio) instead. This brings the Flutter VAD surface in line with the Swift capability (RunAnywhere+VAD.swift).

Changed #

• v2 architecture: All public capability surfaces (RunAnywhere.llm, .stt, .tts, .vad, .vlm, .voice, .embeddings, .tools, .rag, .models, .modelLifecycle, .downloads, .hardware, .solutions, .lora) are now namespaced accessors on the RunAnywhere static entry point, mirroring the Swift/Kotlin/RN/Web SDKs.
• Lifecycle-owned LLM/STT/TTS/VLM loads: Model loading now routes through commons model lifecycle via the generated proto ABIs (rac_*_lifecycle_proto). Per-component handle ownership has moved into C++.
• Public Voice eventStream(): Voice agent surface now exposes a proto-typed eventStream mirroring Swift's RunAnywhere.voice.events.
• Tool-calling session ABIs: Tool-calling now uses session-based proto requests (rac_tool_calling_*_proto) for parity with Swift/Kotlin.
• Plugin loader capability: New RunAnywherePluginLoader surface for registering backend plugins (LlamaCpp/ONNX/QHexRT) from Dart.
• Model lifecycle / registry split: RunAnywhereModels (registry) and RunAnywhereModelLifecycle (load/unload/current) are now separate capability classes.
• Convenience codegen: Added generated *.defaults() factories (VADConfiguration.defaults(), LLMOptions.defaults(), etc.) so callers no longer hand-construct proto messages with magic defaults.

Added #

• QHexRT NPU Backend: Added NPU framework enum support for Qualcomm Hexagon bundles
• RAG Types: Extended RAG type definitions for enhanced retrieval-augmented generation

Changed #

• Updated model type definitions to support NPU framework registration

Added #

• API Configuration: Custom API configuration management for flexible backend routing
• Keychain Store: Secure credential storage capabilities via platform keychain
• Dev Mode: Development configuration support for local testing and debugging

Fixed #

• Parameter Piping: Fixed parameter propagation through SDK layers (#340)
• Network Layer: Resolved authentication and dev config networking issues
• Simulator Scripts: Fixed build scripts for iOS simulator targets
• Android Models: Updated model support handling for Android platform

Changed #

• Updated native binaries (RACommons, RABackendLLAMACPP, RABackendONNX) to latest builds
• Addressed PR #309 review feedback with critical fixes

Fixed #

• iOS: Updated RACommons.xcframework to v0.1.5 with correct symbol visibility
  • The v0.1.4 xcframework had symbols that became local during linking
  • v0.1.5 xcframework properly exports all symbols as global
  • Combined with DynamicLibrary.executable() from 0.15.10, iOS now works correctly

Fixed #

• iOS: Fixed symbol lookup by using DynamicLibrary.executable() instead of DynamicLibrary.process()
  • process() uses dlsym(RTLD_DEFAULT) which only finds GLOBAL symbols
  • executable() can find both global and LOCAL symbols in the main binary
  • With static linkage, xcframework symbols become local - this is the correct fix

Fixed #

• iOS: Added linker flags (partial fix, superseded by 0.15.10)

Important #

• iOS Podfile Configuration Required: Users must configure their Podfile with use_frameworks! :linkage => :static for the SDK to work correctly. See README.md for complete setup instructions.

Added #

• Initial public release on pub.flutter-io.cn
• Core SDK infrastructure with modular backend support
• Event bus for component communication
• Service container for dependency injection
• Native FFI bridge for on-device AI inference
• Audio capture and playback management
• Model download and management system
• Voice session management
• Structured output handling for LLM responses

Features #

• Speech-to-Text (STT) interface
• Text-to-Speech (TTS) interface with system TTS fallback
• Voice Activity Detection (VAD) interface
• LLM text generation interface with streaming support
• Voice agent orchestration
• Secure storage for API keys and credentials

Platforms #

• iOS 13.0+ support
• Android API 24+ support