Layla

New features: - VRM character models can run as animated, context-aware chat backgrounds with expressions and lip-sync during voice calls. - Contextual mini-apps can run inside chats and respond to the active conversation. - Added contextual SDK events for execution context, sentiment changes, thinking, messages, and voice playback. - Custom mini-apps can be pinned to the Welcome Screen or replace the default character. - Local Dream now supports LoRAs alongside compatible safetensor base models. - Updated Local Dream and redesigned the Image Upscaler with additional models and closer integration. - Added controls for which URLs custom mini-apps may access, including fully offline and unrestricted modes. - Added SDK APIs for saving TTS as WAV or MP3, background audio playback, listing image models, and checking the installed Layla version. - Companion Mode v2 adds direct access to enabled Agents. - Voice and text controls can be enabled independently, with a full-screen text chat. - Companion Mode loads its LLM on demand and unloads it after five minutes of inactivity. - The Layla SDK now supports image recognition for mini-apps and characters. - MobileVLM now supports custom on-device image recognition models. - Added support for SDXL MNN image-generation models. Improvements: - Mini-app settings now save automatically. - Mini-apps can reference scripts, styles, and other local files from their own directory. - Mini-app uploads can include selected files and directories instead of the entire app directory. - Custom mini-apps can now execute WebAssembly. - Expanded the default list of allowed CDN URLs. - Console messages can now be viewed inside custom mini-apps. - Unrecognised SDK API calls now return an error instead of remaining pending. - Added SD WebUI and upscaler settings to Local Dream, completing the deprecation of the Stable Diffusion mini-app. - Model benchmark reports include relevant Advanced Settings and can be searched and sorted. - Voice-chat TTS has fewer pauses and queues simultaneous requests more reliably. - Raised the recall threshold for memories created by the current chat. - Switching models now preserves temporary downloads such as mini-app icons. - Image generation can now be stopped before completion, including through the Layla SDK. - Model reasoning is preserved for display in chat history without being sent back to the model. - Scheduled chat notifications now include the character’s image. - Roleplay chats now include a Generate image action and full-screen image viewing. Bug fixes: - Fixed GGUF models stored on an SD card not loading. - Fixed models returning an empty message after using REGEN or reloading. - Fixed LiteRT benchmarks reporting zero for performance statistics. - Fixed duplicate system prompts appearing in chat debug information. - Fixed incorrect warnings that a previous REPL instance was still running. - Fixed tool calling stopping after malformed model output. - Fixed folder renaming not accepting a new name. - Fixed missing character profile images in Long-term Memory. - Fixed MCP tools appearing as null in past conversations. - Fixed chat engine errors when continuing a conversation after sending an image. - Fixed custom image-generation Agents not being triggered. - Fixed logcat collection in Advanced Settings. - Fixed VRM characters not responding to taps during chat. - Fixed the final memory created before leaving a chat not receiving a summary.