cera_ffi 0.5.4

Dart bindings for the Cera on-device inference engine. Runs GGUF LLMs locally through dart:ffi, with no Flutter dependency.

Changelog

0.5.4

Fixed

• Model and Adapter Loading: Added fallback path loaders for LoRA adapters, Silero VAD, and Hybrid PII models when cera is compiled without default features (mmap disabled) on native targets.

Changed

• Version Alignment: Bumped workspace patch version to 0.5.4 in lockstep with cera_ffi_flutter and the underlying native engine crates.

0.5.3

Added

• Streaming Text Chunks: Added generateStreamChunks API for streaming multi-token text chunks, reducing cross-boundary messaging overhead.
• Multimodal Envelopes: Added MultimodalEnvelope, ImageInput, and AudioInput types for structured multimodal prompt delivery.
• Android Download Service: Added support for background model asset downloading via AndroidDownloadService.
• Whisper ASR Integration: Added native bindings for OpenAI Whisper speech recognition models.

0.5.2

Changed

• Version Alignment: Bumped workspace patch version to 0.5.2 in lockstep with cera_ffi_flutter and the underlying native engine crates.

0.5.1

Added

• WebGPU Dense Transformer Support: Direct GPU execution support for dense transformer architectures (llama, qwen2, qwen3, granite) in addition to lfm2/lfm2.5/lfm2moe.
• License Formatting: Cleaned dual Apache-2.0 / MIT license formatting for automated OSI recognition by pana on pub.dev.

0.5.0

Added

• Native Silero VAD v5 engine and bindings: Added FfiSileroVad, FfiVadSampleRate, FfiVadIterator, FfiVadConfig, and sileroVadDefaultConfig for real-time speech activity detection, streaming audio chunk processing, and speech segment timestamping.
• Hugging Face model repository support: Direct loading from Hugging Face model repository specs and URLs with optional on-the-fly streaming quantization.
• Multimodal audio & vision enhancements: WebGPU Depthformer acceleration, 4 voice modes (SpeechToText, VoiceChat, TextToSpeech, TextOnly), microphone audio input, silence trimming, and vision encoder ViT optimizations.

0.4.0

First release. The Dart bindings previously lived inside the cera_ffi_flutter package; they were split out here so that they can be used without Flutter at all.

Added

• Cera, a portable asynchronous API that runs on every target including the web. One surface over two transports: the Rust async runtime natively, a Web Worker running cera-wasm in a browser. Loading, chat templating, tokenizing and streaming generation, with generations serialized against one KV cache. The generated bindings stay synchronous and native-only; Cera exists because a browser cannot offer a synchronous generate at any price.
• Web inference, on WebGPU where the browser has it and on a wasm CPU build where it does not (58 tok/s against 1.4 tok/s on the same machine and model). dart run cera_ffi:install_web puts the runtime in an app's web/; no COOP/COEP headers are required.
• CeraEngine.fromPathAsync and fromBytesAsync, so loading a model no longer blocks the calling isolate.
• dart:ffi bindings for the Cera inference engine, generated from the compiled cera-ffi cdylib by a vendored uniffi-bindgen-dart and committed under lib/src/generated/.
• CeraLibrary, a platform-aware loader: an explicit path, then CERA_FFI_LIB, then the platform's normal search path.
• Plain-Dart examples under example/: chat, streaming, async, and download progress.

Notes

• This package resolves under plain dart pub get, which is the reason it exists. pub will not publish a package declaring flutter.plugin.platforms without a Flutter SDK constraint, and declaring that constraint makes dart pub get refuse the package outright. A Flutter plugin therefore cannot also be a plain-Dart package, so the bindings live here and cera_ffi_flutter depends on them.
• No native library ships with this package. Flutter apps get one from cera_ffi_flutter; everyone else points CERA_FFI_LIB at a cera-ffi build. That build needs the ffi-buffer cargo feature, without which every call fails at dlsym.
• The bindings are still exported conditionally, and the web branch is still a generated stub with the same API and no dart:ffi: data types are real, engine entry points throw UnsupportedError. That is what lets a multi-platform app build at all, since an unconditional dart:ffi import fails the whole build rather than one branch.