AIRI v0.12.0-beta.1 release notes

This beta adds new character formats, stronger desktop controls, and major reliability work across chat, speech recognition, and provider settings.

@ricemaster1 made their first contribution in https://github.com/moeru-ai/airi/pull/2016
@juyua9 made their first contribution in https://github.com/moeru-ai/airi/pull/1803
@wuyua9 made their first contribution in https://github.com/moeru-ai/airi/pull/1911
@Anfi1 made their first contribution in https://github.com/moeru-ai/airi/pull/2072
@ShiroKSH made their first contribution in https://github.com/moeru-ai/airi/pull/2054
@kobihikri made their first contribution in https://github.com/moeru-ai/airi/pull/2149
@vuyua9 made their first contribution in https://github.com/moeru-ai/airi/pull/1871
@LauraGPT made their first contribution in https://github.com/moeru-ai/airi/pull/2207
@MilkyWeighW made their first contribution in https://github.com/moeru-ai/airi/pull/2205
@ztcools made their first contribution in https://github.com/moeru-ai/airi/pull/2178
@xianjianlf2 made their first contribution in https://github.com/moeru-ai/airi/pull/2280
@yzxcj797 made their first contribution in https://github.com/moeru-ai/airi/pull/2300
@ppdbxdawj made their first contribution in https://github.com/moeru-ai/airi/pull/2321

Desktop ver. now supports MMD models. You can import motions and adjust physics, gaze, lip sync, expressions, and materials.
Desktop ver. now supports Tachie archives that switch complete character images to match emotions.
VRM characters now react when you click their heads, arms, hands, or feet.
Live2D imports now support CJK paths and older filename encodings.
If a Live2D model fails to render, the Stage now shows a clear error and remains usable.

You can now mute voice output across sessions or stop all current speech from the Controls Island.
The new Hearing playground lets you monitor the active speech-recognition pipeline and review each completed speech segment.
Speech recognition now keeps transcript corrections and handles voice activity, pauses, and streaming audio more reliably.
Long conversations use less rendering work, so chat remains responsive as the history grows.
Tool messages now remain available across chat turns.
Streamed messages no longer flash, and tool-call progress stays synchronized.
Chat math now uses consistent syntax for predictable display.

The Controls Island now moves to the nearest window corner after you reposition the desktop window.
Connection status now lives inside the Controls Island, and tray-based window moves use animations.
Desktop pairing settings and QR codes now stay synchronized.
Desktop ver. can relay conversations to a connected Minecraft bot and read supported bot speech aloud.
AIRI now starts on Linux ARM64. Flatpak installs also include desktop and AppStream metadata.
Mobile ver. now has a permission panel, clearer onboarding, and Android microphone selection.
On mobile browsers, the software keyboard no longer hides or shifts the chat input unexpectedly.

You can enable Thinking for supported Consciousness models from one model option.
DeepSeek now applies its reasoning mode correctly.
Provider validation status now persists, and each provider uses its own validation rules.
Tool use works with more providers because AIRI adapts tool schemas to each provider.
Volcengine Coding Plan now lists current models with localized descriptions.
Steam sign-in and account linking are now available.
Hosted authentication now supports Apple sign-in and native Apple ID tokens.
Account deletion now revokes linked social authorizations.
Hosted model routing can fall back across configured provider groups when one route is unavailable.

A provider settings loop could leave AIRI loading forever. The startup and synchronization path is now stable.
Desktop chat creation, switching, deletion, and import refresh now work reliably.
Chat and sign-in state now remain consistent across Electron windows.
AIRI Card drafts, active sessions, avatar Models, and runtime data now remain intact during edits and activation.
Whisper worker errors now stop promptly instead of leaving transcription stuck.
Stopping Kokoro speech no longer destroys the worker needed for the next response.
Safari no longer blocks the authentication interface while it loads code chunks.

The CCv3 codec now accepts older and newer cards while preserving unknown fields for forward compatibility.
Hosted chat WebSockets now expose versioned Eventa contracts. Version 2 authenticates after connection and enforces payload limits.
The desktop plugin SDK now preserves nullable enum schemas correctly.
MMD and Tachie renderer packages expose reusable scene components and runtime controls.
Hosted configuration values can reload from PostgreSQL without a service restart.

Audio tests can now drive the full input and output pipeline with deterministic microphone fixtures.
Vitest browser projects now run headlessly in CI.
Vishot now lives in its public repository. AIRI also has updated capture and pull-request skills.
Polaroid captures now include imported Live2D models.
The new i18n glossary and Crowdin workflow keep product terms consistent across translations.
Release workflows now support Steam desktop deployment, image provenance, SBOM attestations, and automated README artifact updates.
The documentation site now includes Korean content and smoother search, theme, and page transitions.

Run the Resource API and Auth as separate services.
Set AUTH_SERVER_URL, AUTH_SERVER_INTERNAL_URL, PUBLIC_URL, and RESOURCE_SERVER_URL for the new service boundary.
On Railway, set RATE_LIMIT_TRUSTED_PROXY=railway only for services that receive Railway proxy traffic.
Do not run shared database migrations from Auth. The Resource API remains the only migration owner.
Before you build or run Godot tests, install the .NET 10 SDK.

Full Changelog: https://github.com/moeru-ai/airi/compare/v0.11.3...v0.12.0-beta.1

AIRI v0.12.0-beta.1