kiln_flutter 0.5.2

On-device AI for Flutter on Android: chat, decision, embedding, speech-to-text and text-to-speech models, with honest stats.

0.5.2

• Shorter package description (pub.dev asks for 60 to 180 characters) and dart format on the library and example. No change to the code's behaviour.

0.5.1

• Phones without the dot-product instruction no longer crash. The plugin now also ships libkiln_armv8.so, built for plain Armv8.0 (armv8-a). native.dart loads it when /proc/cpuinfo lacks asimddp, asimdhp or fphp; other phones load libkiln.so as before. Verified on an OPPO CPH2819 (Snapdragon 680, 3.76 GB RAM): the example app and the Hishab app load libkiln_armv8.so and answer. Not yet verified: libkiln.so on a Pixel 8 or S26 with this version.
• cancel() on a generation, transcription or speech now also stops it when it is called before the worker isolate has started the call (right after chat(), for example). Such a cancel used to be lost and the call ran to the end. Each call has a cancel flag in native memory that the worker checks before starting and at every piece.
• With kiln's core change, a cancel during prompt rendering is no longer lost either.

0.5.0

Text to speech. Verified on a Pixel 8: 26 integration tests (4 skipped: the speed and media groups, which passed at 0.3.1), and the example app's new Speak screen. Android arm64 only, as before.

• KilnContext.speak(text, {voice, language, maxFrames, seed}): says the text in the voice of a short recording (KilnMedia, a file or its bytes) and streams the audio as KilnAudioPieces (mono floats, 24 kHz) about once a second; KilnSpeech.wav collects them into a wav file's bytes, .stats gives the length, the time and how many times faster than real time. pcmToWav is public for your own pieces.
• KilnProjectorInfo.generatesAudio and .sampleRate: a text-to-speech projector.
• Tested with Pocket TTS (Kyutai, English, 159 MB + 60 MB projector; CC-BY-4.0): 2.7 s of speech in 3.4 to 4.1 s (0.7 to 0.9x real time; about a second of it is encoding the voice) on the Pixel 8. The example app's Speak screen bundles a sample voice, records your own, plays the result with audioplayers, and has Pocket TTS in its download list.
• The integration test checks the audio by having Parakeet transcribe it. Through the plugin the sentence took 4 to 8 s on a Pixel 8 that had been running tests for hours (0.4 to 0.6x real time), 5.3 s in the example app; the C API on a rested phone does it in 3.4 to 4.1 s. Let a piece or two arrive before playing, or keep sentences short.

0.4.0

Published 2026-10-13. Speech to text. Verified on a Pixel 8 (Tensor G3): 25 integration tests (4 skipped: the speed and media groups, which passed at 0.3.1), and the example app's new Transcribe screen (record, or pick a file). Android arm64 only, as before.

• KilnContext.transcribe(KilnMedia audio, {language, prompt, maxTokens}): writes down a wav, mp3 or flac clip (a file or its bytes) and streams the words; KilnTranscription.text collects them, .stats gives the clip length, the time taken, how many times faster than real time, and the language the model heard. cancel() stops it.
• Three model families. Parakeet TDT 0.6B v3 (669 MB) and Whisper base (60 MB) come as one .bin file each in whisper.cpp's format; loadModel takes them like a GGUF, KilnModelInfo.isSpeechModel is true, and they need no projector. Qwen3-ASR 0.6B is a GGUF plus its projector (loadProjector), like the embedding media models, and isSpeechModel is false for it.
• Speed on the Pixel 8 for a 3-second clip: 0.3 s with Parakeet (0.8 to 1.5 s once the phone is hot), 3.0 s with Whisper base, 3.8 s with Qwen3-ASR (slower than the clip; its encoder takes 2.5 s on that CPU). Through the plugin: Parakeet 1.2 s on a phone that had been running tests for an hour, Whisper base 7 s.
• Example app: a Transcribe screen with a microphone button (records a 16 kHz wav with the record package, needs RECORD_AUDIO) and a file picker; Parakeet and Whisper base in the download list; .bin files in the model list and the paste-a-link box.
• KilnModelInfo.toString() says "speech model"; KilnMedia is shared by embedMedia and transcribe.

0.3.1

• Phones whose chip has no little cores now use more of it. On a Galaxy S26 Ultra (Snapdragon 8 Elite Gen 5) kiln ran on 2 of 8 cores; it now runs 6 threads. Chat with LFM2.5 350M and Qwen3.5 0.8B is about 20% faster under sustained load, and a decision or an embedding takes 35 to 45% less time. Phones with little cores, such as the Pixel 8, are unchanged. KilnDeviceInfo.cpuFastCores reports the new choice.
• Verified on a Galaxy S26 Ultra (26 integration tests and the native test suite) and on a Pixel 8 (native test suite; its integration test was last run at 0.3.0). No API change. Added after the release: the 26 integration tests also pass on the Pixel 8 at 0.3.1.

0.3.0

Embedding models. Verified on a Pixel 8 (Tensor G3): 23 integration tests, including images and audio. Android arm64 only, as before.

• KilnContext.embed: turn a text into a vector with an embedding model, for search by meaning. kind (query or document) adds the prefix the model was trained with, dimensions keeps a shorter vector, normalize: false returns the raw one.
• KilnModel.loadProjector and KilnContext.embedMedia: images and audio, alone or mixed with text, into the same space as text. KilnMedia.file or KilnMedia.bytes.
• KilnEmbedding holds the vector (Float32List) and the stats; similarity and KilnEmbedding.cosine compare two.
• KilnModelInfo.isEmbeddingModel and nEmbd.
• chat and decide on an embedding model throw KILN_ERR_NOT_SUPPORTED, as embed does on the other kinds.
• The native library now includes llama.cpp's image and audio encoders (libkiln.so grows from about 6 MB to about 7 MB).
• Example app: a Search screen for embedding models, with a picker for photos and audio.

0.2.0

First version on pub.dev. Android arm64 only; iOS is in progress and not part of the package. Verified on a Pixel 8 (Tensor G3): 16 integration tests. Not yet tried on Snapdragon or MediaTek phones or on devices with 4 GB of RAM.

• chat(maxReasoningTokens:): thinking budget. After that many reasoning tokens kiln closes the think block and the model answers. KilnGenStats.reasoningTruncated reports when it was used.
• KilnContext.decide: ask a decision model (LFM2 d1-omni) a yes / no, choice or score question about a text state and get a probability per option. KilnModelInfo.isDecisionModel tells the two kinds of model apart.
• Hybrid models (Qwen3.5, LFM2) reuse the KV cache across turns through state checkpoints.

0.1.0

• First release. Android arm64, llama.cpp backend, streaming chat with reasoning channel, KV prefix reuse, cancel, device probe, model info and generation stats.