NewYour coding agent can read the release notes before it upgrades.Set up the MCP server →
pub.dev
pub.dev, #2341 most downloaded on pub.dev
On-device AI for Flutter: run Gemma and other LLMs and embeddings on Android, iOS, Web, macOS, Windows and Linux, and speech on all but Web. Pluggable engines, no server.
Latest stable
2.6.0
Last release 14 days ago
08 Nov 2026
Release timing varies
gaps range from 8 days to 3 months
Nearly every release is documented
notes for 22 of 22 stable releases
Nothing withdrawn
no release was ever pulled
4 months old
23 releases · first in 2026
Drag across the ruler to compare, or tap a release to open it.
createSession, createChat, session and chat are deprecated until 3.0.0.
createSession, createChat, session and chat are deprecated until 3.0.0.ContextLimitException reports prompts that exceed the context window.InferenceModel; remove OptionScoringModel (#769).createModel stays deprecated until 3.0.0, not 2.6.0 (#768).
createModel stays deprecated until 3.0.0, not 2.6.0 (#768).openSession/openChat (onnx 0.8.0) and memory diagnostics on Windows.Your coding agent can read these notes before it upgrades. Set up the MCP server →
One column per month.
uninstallModel removes a directory (ORT-GenAI) model by its name (#730).
uninstallModel removes a directory (ORT-GenAI) model by its name (#730).uninstallEmbedder, uninstallStt and uninstallTts close the model loaded from those files (#735).installModelFromCatalog, fromCatalog, and expectedSize checks on network installs (#734).installTts().withProgress counts bytes across the bundle, not finished files (#751).ModelType.gemma4 on .bin/.tflite: Gemma 4's own turn markers, not Gemma 3's (#705).
ModelType.gemma4 on .bin/.tflite: Gemma 4's own turn markers, not Gemma 3's (#705).uninstallModel closes the model built from that file; getModelPath follows the file's record (#723).
uninstallModel closes the model built from that file; getModelPath follows the file's record (#723).InferenceModel.createDecider reads LiteTune 0.6.0 files; one scored fresh_engine is refused by name.
InferenceModel.createDecider reads LiteTune 0.6.0 files; one scored fresh_engine is refused by name.Android: the manifest declares MediaTek's NeuroPilot libraries for the MediaTek NPU.
post_install block; the load error names the fix (#689).flutter_edge_ai_llama 0.0.3.flutter_edge_ai_litertlm 1.17.0.Breaking: remove the members deprecated for 2.5.0 (listed in MIGRATION.md).
FlutterEdgeAi.openModel: several caller-owned models at once; FlutterEdgeAiPlugin.createModel deprecated.InferenceModel.createDecider scores options on a model trained on the decision, Noul and Score too.honouredBackends, the loaded backend reported, .gguf installs without a tokenizer.Android NPU: one vendor's dispatch per load; the manifest declares Google Tensor's runtime libraries.
generateChatResponseWithTools: a failed feed no longer masks the tool's error; unencodable results are answered.Experimental Decider API: pick one of N options or abstain (EmbeddingModel.createDecider).
Decider API: pick one of N options or abstain (EmbeddingModel.createDecider).implements InferenceModel add createDecider and two getters; extends is unaffected.getActiveModel(multimodalOrder:): per-model order of images, audio and text.Message.toolCall(calls:), toolCallId, FunctionCallResponse.id, toolCallsOf.InferenceModel.activeVisionBackend/activeAudioBackend; an unset encoder backend is the engine's choice.Docs: README and shipped skills checked against the code; models page linked.
The active model keeps its LoRA across a restart; ModelFileType.gguf (#649).
ModelFileType.gguf (#649).Combined text, image and audio inputs: EmbeddingInput(text:, images:, audio:).
EmbeddingInput(text:, images:, audio:).Breaking: remove deprecated 1.x/2.0 members (listed in MIGRATION.md).
embed, EmbeddingInput); see MIGRATION.md for minimum versions.EmbeddingInputTooLongException instead of truncating.lib/src; engines import package:flutter_edge_ai/engine.dart.getActiveModel waits for a closing model to finish, then builds the next one.
getActiveModel waits for a closing model to finish, then builds the next one.getActiveModel reloads the model.NoActiveModelError, a StateError callers can tell apart.implements InferenceModel: add isClosed; extends inherits the default.fromFile is active again after a restart.fromFile files.Breaking: unset sampling values come from the model file, then its family, not greedy topK: 1 (#572).
topK: 1 (#572).Closing an embedder mid-batch no longer leaks its native model.
Allow code_assets 2.x, which other native plugins now require (#603).
code_assets 2.x, which other native plugins now require (#603).A crash while saving the active inference, STT or TTS model no longer leaves a mix.
Breaking: isThinking is enableThinking; ModelRuntimeDefaults.isThinking is thinkingDeclared.
isThinking is enableThinking; ModelRuntimeDefaults.isThinking is thinkingDeclared.ModelType.qwen35; thought-channel reasoning is a ThinkingResponse for every type; no /no_think on audio.installTts() also takes fromAsset, fromFile and fromBundled; install() throws on web.fromAsset installs on desktop.flutter_gemma becomes Flutter Edge AI: every package renamed to flutter_edge_ai* by @DenisovAV in https://github.com/DenisovAV/flutter_edge_ai/pull/56
Full Changelog: https://github.com/DenisovAV/flutter_edge_ai/compare/v1.11.4...v2.0.0
flutter_edge_ai_rag.ModelFileManager.setActiveModel; use ensureModelReadyFromSpec.flutter_gemma name aliases; dart fix --apply still migrates.The codelabs carry the litert_embeddings.js that flutter_gemma_litertlm 1.8.4 ships by @DenisovAV in https://github.com/DenisovAV/flutter_edge_ai/pull
Full Changelog: https://github.com/DenisovAV/flutter_edge_ai/compare/v1.11.0...v1.11.4
flutter_gemma; dart fix --apply migrates, old names deprecated until 2.0.0.Nothing published for this version