- fixed a crash after music generation

- fixed a crash using MobileVLM - fixed missing translations in the new settings page

- fixed a bug where Layla Cloud asks you to login repeatedly - fixed a bug where model benchmarks were not showing progress

New features: - On-device music generation with ACE-Step 1.5, including optional lyrics and controls for prompts, models, generation steps, and more. - ACE-Step music-generation APIs for mini-app creators, including lower-level lm, synth, understand, and vae APIs. - SillyTavern preset support: manage presets in a new mini-app and attach them while creating or editing characters. - Experimental model recommendations based on benchmark results from users with similar devices. - Background tasks for mini-apps through standalone task.js scripts, restoring background message scheduling for Dream and other mini-apps. - A redesigned Task Manager for viewing and managing active, failed, disabled, and unsupported background tasks. - New file-management APIs for mini-apps, including directory listing, file/directory deletion, and relative paths for saved and read files. - Use generated images directly in the character-creation flow without saving them to your phone first. - Concurrent SDK calls, such as generating text-to-speech while an LLM response streams. Improvements: - Updated llama.cpp and ACE-Step OpenCL kernels for better compatibility with older mobile GPUs. - Redesigned the Settings page, Download Manager mini-app, and Create Character screen. - Added a Restart Later option so multiple restart-dependent settings can be changed before restarting Layla. - Added a clear-filters option to Character Selection. - Improved model-download stability. - Improved the display of chats auto-translated by the Offline Translator mini-app. - Image-generation prompts can now be hidden from the chat display. - Updated llama.cpp to the latest commit. Bug fixes: - Fixed a crash when selecting a colour. - Fixed a crash caused by line breaks in additional llama.cpp arguments. - Fixed input suffixes appearing after the system prompt. - Fixed reference-document attachments not showing as attached during chats. - Fixed character greetings not being sent to the model. - Fixed a black background on some phones when Layla is set as the default assistant. - Fixed chat message statuses and other special message content not displaying correctly when loading conversation history.

New features: - On-device music generation with ACE-Step 1.5, including optional lyrics and controls for prompts, models, generation steps, and more. - ACE-Step music-generation APIs for mini-app creators, including lower-level lm, synth, understand, and vae APIs. - SillyTavern preset support: manage presets in a new mini-app and attach them while creating or editing characters. - Experimental model recommendations based on benchmark results from users with similar devices. - Background tasks for mini-apps through standalone task.js scripts, restoring background message scheduling for Dream and other mini-apps. - A redesigned Task Manager for viewing and managing active, failed, disabled, and unsupported background tasks. - New file-management APIs for mini-apps, including directory listing, file/directory deletion, and relative paths for saved and read files. - Use generated images directly in the character-creation flow without saving them to your phone first. - Concurrent SDK calls, such as generating text-to-speech while an LLM response streams. Improvements: - Updated llama.cpp and ACE-Step OpenCL kernels for better compatibility with older mobile GPUs. - Redesigned the Settings page, Download Manager mini-app, and Create Character screen. - Added a Restart Later option so multiple restart-dependent settings can be changed before restarting Layla. - Added a clear-filters option to Character Selection. - Improved model-download stability. - Improved the display of chats auto-translated by the Offline Translator mini-app. - Image-generation prompts can now be hidden from the chat display. - Updated llama.cpp to the latest commit. Bug fixes: - Fixed a crash when selecting a colour. - Fixed a crash caused by line breaks in additional llama.cpp arguments. - Fixed input suffixes appearing after the system prompt. - Fixed reference-document attachments not showing as attached during chats. - Fixed character greetings not being sent to the model. - Fixed a black background on some phones when Layla is set as the default assistant. - Fixed chat message statuses and other special message content not displaying correctly when loading conversation history.

New features: - Layla Python SDK: a new Python interface for agents that runs outside the UI — run inference, start speech, generate images, and drive an automated web browser - Web browser automation: your character can visit websites, open links, look up news, and grab images on its own - Perchance Image Generation Agent: opens a browser, generates an image, and shows it in your chat - On-device image generation through LiteRT (Bonsai 4B, Flux2.Klein 4B, and Z-Image-Turbo) Improvements: - Redesigned the Character Selection, chat header, Chat Info, and settings screens - Rebuilt the Inference Settings and Advanced Settings pages for easier preset management - Stable Diffusion model settings are now included in backups - Layla now warns before leaving a custom mini-app when you press Back Bug fixes: - Fixed empty model responses when quickly switching between chats - Fixed regenerating a message after editing or deleting the previous one - Fixed Layla locking every time you started a chat from history (unlock now lasts 10 minutes) - Fixed the back button only minimising Layla during a chat - Fixed zoom not working in the image gallery - Fixed the Perchance agent not working in Roleplay - Fixed DuckDuckGo web search

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.

- fixed a bug where Kokoro TTS was crashing - fixed a bug where Chat Debug Info cannot be scrolled - fixed a bug where Long-term memory knowledge graphs were not being displayed on iOS

Layla 7.0 adds the Layla SDK and user-made mini-apps, rebuilds long-term memory around a new on-device pipeline, and brings a new set of voices. This release also includes every fix from 7.0.1 through 7.0.4. NEW - Layla SDK: build your own mini-apps as a single HTML file, a zip folder, or a self-hosted site. Mini-apps can use the LLM, characters, image generation, agents, memory and voice. - Browse Apps: install mini-apps made by other users, upload your own to Layla Cloud, manage versions in the Mini-Apps Manager, and share mini-app links. - Agent skill for Codex, Claude Code and OpenCode, plus a mini-app template for larger projects. - Real-time long-term memory, processed as you chat rather than ingested overnight in the background. - New default memory pipeline using a custom-trained model that builds summaries and knowledge graphs together. - Temporal knowledge graphs, with a full-screen view. - Dream rebuilt as a downloadable mini-app, with evolving character impressions, self-reflection and smarter conversation continuation. - Kokoro v1.0 voices, including multilingual options. - File Manager can import arbitrary files, including assets and draft models, and now shows image previews. - Expanded SDK APIs: chat sessions and history, sentiment detection, file save and read, image-to-image, TTS voice listing and generation, message scheduling, persona lookup, top-memory recall, and getting or setting inference engines. IMPROVED - Deleting or regenerating a message now updates or removes the memory it created. - Memory recall skips the most recent same-session memories to avoid feedback loops. - A mini-app can replace the default memory processor through the SDK. - SDK chat responses return reasoning and message content in separate fields. - Faster character selection in very large libraries. - Faster Browse Apps loading, with mini-app icons cached for offline use. - Better handling of large mini-apps that bundle image and audio into their HTML. - File Manager shows total folder size and is faster and more stable with many files. - Backups now include character folders and sorting. - Reference documents and agents are easier to select during character creation. - Sentiment detection aggregates across sentences and normalises pronouns. - Importing a mini-app that already exists asks for confirmation and overwrites it instead of blocking. - Benchmark mini-app redesigned to match the current visual style. FIXED - Long-term memory generation getting stuck while building knowledge graphs. - Knowledge graphs opening to a blank page. - Long-term memory recording messages for characters with LTM turned off. - Removing long-term memory with the References mini-app installed failing with an unknown error. - Scheduled future messages appearing on the chat history page. - Dreams skipping scheduling when using cloud models. - Image-to-image not applying the prompt prefix. - Image generation starting while the model was still reasoning about the image prompt. - Mini-app image generation reading default settings instead of saved settings. - Custom sampler presets not saving additional llama.cpp arguments. - Speak Responses reporting that TTS was not initialised. - Layla Cloud voices not stopping when requested, and the SDK stop-speaking call doing nothing. - PocketTTS not working through the SDK. - Voice chat minimising text with the setting turned off. - Mini-apps sending a system prompt Layla did not use. - Chat history appearing when swiping left in a mini-app. - Importing a Roleplay from a RoleplayPNG not saving its characters. - Roleplay not loading long conversation histories properly. - Auto-translated messages not showing. - Prompts being processed again unnecessarily. - A crash when loading a very large conversation history. Layla runs on your device: no account, no subscription, no data leaving the phone.

The Roleplay mini-app has been completely redesigned! Choose from 50+ templates to start your roleplay sessions: romance, adventure, isekai, comedy, interview practice, language learning, and more! - Roleplay now supports local models (GGUF), connecting to your PC/OpenAI, Layla Server, or Layla Cloud. - edit/create your own scenarios with an all new editor - create temporary characters directly within the Roleplay mini-app, or attach your existing characters to the scenario. (Your temporary characters will be automatically put into a folder with the name of your Roleplay in the character list) - Visual Novel mode: centres the character image as they speak/act - OOC Instructions: you can add out-of-character instructions at any time during the roleplay, useful to keep the AI on track - TTS: text-to-speech for individual characters in their own voices (configuration in the character edit) - Roleplay supports custom agents and image generation - Roleplay automatically switches LLMs based on your saved inference engines depending on who is talking, making it finally possible to have multiple LLMs chat with each other and see how they interact! Improvements: - improved long-term memory summarisation speed - added Gemma4 (thinking) preset prompt - redesigned the Saved Inference Engine section to easier filter by attached characters - improved TTS consistency by removing unnecessary stopping mid-sentences - updated Local Dream to add image editing features - added new schedulers (Kerras versions) in Local Dream Bug fixes: - fixed a bug where the model selector disappears after searching for available models when configuring OpenAI API - fixed a bug where chat bubble styles stops showing when performance counter is enabled - fixed a bug where changing the prompt but not the seed in image generation does not generate a new image - fixed a bug where "Speak Responses" in chat actions cannot be toggled off - fixed a bug where "Push to talk" had stopped working

- fixed a bug where custom vision models were not being detected correctly - fixed a bug where translation quick actions was not available - fixed a bug where Stable Diffusion models were stuck on generating

Layla supports one-shot Voice Cloning with PocketTTS mini-app! Create your own voices by recording or uploading an audio sample. Layla will create a custom voice out of it for which you can use for your characters! You can also set your custom voice as the the global TTS voice for Layla, or configure it individually for each character. Record audio right in Layla to create custom voices: you can record an audio snippet and use it as a base to clone a voice Voice Cloning tips: - PocketTTS is very sensitive to background noise. Record your audio in a quiet room or use an online editor to remove background noise to get better results - only supports 16-bit PCM wav files, use an online converter if you need to convert mp3's etc. - try not to have blanks/silences in the sample, PocketTTS will try to learn that and generate lots of silences - exaggerated voices work better New features: - added ability to configure chat-bubble colours in UI settings Improvements: - gallery of generated images in Layla are now backed up - flipped the user/AI position of chat bubble styles to match the standard for other social media/chat apps - image gallery now blocks other apps from taking a screenshot or recording of it's contents, and will also automatically hide its contents when Layla goes into the background - image gallery in Layla allows multi-selection for deleting images - improved TTS latency by implementing audio streaming - you can re-order characters inside folders and drag them outside to remove them from the folder - improved character impression generation by limiting the max output and cutting off incomplete sentences - Stable Diffusion prompts are saved in generated images gallery Bug fixes: - fixed a bug where Layla Cloud image generation fails to read the generated images after a few attempts - fixed a bug where default Automatic1111 json body stopped working - fixed a bug where setting a max response length does not trim incomplete sentences - fixed a bug where previewing native TTS using Google/iOS was not using the selected voice properly in other languages - fixed a bug where sometimes asterisks were not being skipped in TTS when skip asterisk setting on

Improvements: - advanced settings allows adjusting LLM and VLM hardware acceleration separately, allowing you to put the LLM on the CPU while the VLM on the GPU - supports Gemma-4 tool calling agents in both GGUF - allow setting additional llama.cpp command-line arguments in Advanced Settings - additional llama.cpp cmd arguments can be saved as Advanced Settings presets Bug fixes: - fixed a few "Failed to eval" bugs on long context usage - fixed a few UI issues that happens when the screen is horizontal - fixed a bug where viewing a memory that is too long will soft-lock you without any way to exit the chat - LTM now removes think tags from Gemma 4 if used as summariser

New features: - added new Choose Your Own Adventure characters - supports the latest Gemma-4 model family - added German translations Improvements: - added ability to send progress messages during python execution, you can view this progress in the Agent execution card to give you more insight into what the agent is doing - Layla Python can now generate files - new Python agents will automatically install needed dependencies when added for the first time - you can configure long-term memory to use a saved inference engine as summariser - added UI setting to display chats in a bubble-like interface instead (similar to Facebook Messenger or WhatApp etc.) - preset characters are now organised into folders Bug fixes: - fixed a bug where sometimes character expressions will not be saved - fixed a bug where some memories cannot be processed

New features: - added new Choose Your Own Adventure characters - supports the latest Gemma-4 model family - added German translations Improvements: - added ability to send progress messages during python execution, you can view this progress in the Agent execution card to give you more insight into what the agent is doing - Layla Python can now generate files - new Python agents will automatically install needed dependencies when added for the first time - you can configure long-term memory to use a saved inference engine as summariser - added UI setting to display chats in a bubble-like interface instead (similar to Facebook Messenger or WhatApp etc.) - preset characters are now organised into folders Bug fixes: - fixed a bug where sometimes character expressions will not be saved - fixed a bug where some memories cannot be processed

New features: This update fleshes out Layla's Agent Framework. **Agent System Prompts** Agents now have their own system prompts. These will update your character's system prompt when they are activated. For example, you don't need to update both your system prompt and the agent for status cards; the system prompt to guide the model into generating status cards will automatically be injected when you attach the corresponding agent. **Introducing Short-term Memory** Why do we need a short-term memory when there is the LLM context itself? The main difference lies in the fact that Layla's short-term memory is **structured**. Both the LLM context and LTM is unstructured, and lookup is done purely via heuristics (via the Attention layer in LLMs, and embeddings in LTM). The short-term memory is structured a layer that lives between the LLM's context and Layla's LTM. Information flows from the LLM's context -> short-term memory -> LTM (as demonstrated by one of the example agents below). This can be updated by both the LLM and Layla's Agents. Because short-term memory is structured, you can use code and logic to manipulate the contents, not just via prompts. This opens up the possibility to *program* information flow. An example could be writing an agent that updates a character's Hit-Points or inventory based on triggering conditions. This layer will also serve as a sort of *RAM* for all agents. Any agent can read/write into Layla's short-term memory, allowing a horizontal layer of communication between agents (work-in-progress). This is now possible because... **Layla supports Python** Layla now comes with her own embedded Python interpreter as one of the mini-apps. Install it to enable writing and executing arbitrary python scripts directly in your phone! You can install packages via `pip`, and Layla comes with a small console and code area to run your scripts. All agents in Layla can now execute python scripts. Which means you can program arbitrary logic within an agent. You can download stuff from the internet, scrape the web, generate images, solve math problems, even program small games within Layla's Agentic Framework. Improvements: - added setting to disable system prompt in Inference Settings -> Custom Prompts: this will disable all character and user information. This is useful if you want to chat with the "raw model" without any instructions. Note this also stops most apps in Layla from working properly. - redesigned the model selection area, now with the ability to export your imported models - Claude API now allows sending images in chat - added Japanese and Vietnamese translations - added ability to organise characters into folders - you can tap on "memories" in your character card to immediately view their memories in the LTM app - added small UI sounds to notify you when Layla is listening or finished speaking - allow adding custom JSON fields to OpenAI endpoints Bug fixes: - fixed a bug where long prompts will crash LTM ingestion process with a custom LLM - fixed bug where Layla as default assistant on phone does not work with multiple inference engines attached to the character - fixed bug where setting default character was not working with using Layla as default assistant on phone - fixed a bug where custom SD models cannot be imported via the "Already downloaded" link

Improvements: - improved download stability of new model downloads in the welcome screen and stable diffusion model downloads - added TTS setting for a global default voice - added the ability to given your character/LLM a custom instruction during chatting in Chat Actions - added UI setting to control text area expansion in voice chat - think content is removed when reloading the chat (a compromise between removing think content on every message causing a reload, which is too slow for mobile) - added "copy" character button for user created characters - TTS now skips speaking unpronounceable characters (such as ASCII art, tables etc.) - added a button in Diagnostics popup that force restarts Layla Bug fixes: - fixed a bug where Dreams keep scheduling messages for the same character more than once - fixed a bug where blockquotes were not being rendered correctly - fixed a bug where Layla character cannot be duplicated - fixed a bug where different model options in Inference Settings cannot be scrolled sometimes - fixed a bug where "Speak Responses" chat actions were not working - fixed a bug where searching chat history does not search through all chat histories - fixed a bug where GPT-OSS model was not generating responses - fixed a bug where if you cancel voice chat before it initialises the UI dark cover is gone - fixed a bug where edited memories are still using old embeddings, causing recalled content to be different to queried content - fixed a bug where reloading chat sometimes do not read from cached session and instead loads everything from scratch - fixed prompt format for GPT-OSS - fixed a bug where Layla as your phone's default assistant was not working properly

Improvements: - improved download stability of new model downloads in the welcome screen and stable diffusion model downloads - added TTS setting for a global default voice - added the ability to given your character/LLM a custom instruction during chatting in Chat Actions - added UI setting to control text area expansion in voice chat - think content is removed when reloading the chat (a compromise between removing think content on every message causing a reload, which is too slow for mobile) - added "copy" character button for user created characters - TTS now skips speaking unpronounceable characters (such as ASCII art, tables etc.) - added a button in Diagnostics popup that force restarts Layla Bug fixes: - fixed a bug where Dreams keep scheduling messages for the same character more than once - fixed a bug where blockquotes were not being rendered correctly - fixed a bug where Layla character cannot be duplicated - fixed a bug where different model options in Inference Settings cannot be scrolled sometimes - fixed a bug where "Speak Responses" chat actions were not working - fixed a bug where searching chat history does not search through all chat histories - fixed a bug where GPT-OSS model was not generating responses - fixed a bug where if you cancel voice chat before it initialises the UI dark cover is gone - fixed a bug where edited memories are still using old embeddings, causing recalled content to be different to queried content - fixed a bug where reloading chat sometimes do not read from cached session and instead loads everything from scratch - fixed prompt format for GPT-OSS - fixed a bug where Layla as your phone's default assistant was not working properly

- fixed a bug with llama.cpp fix params setting

New features: - redesigned character selection page to give you detailed information of your character before starting the chat - added ability to save custom sampler settings in the Advanced Settings screen - added ability to attach saved sampler presets as part of your custom Inference Engine Improvements - supports sending images to OpenAI compatible API inference settings Bug fixes: - fixed a bug where some app data was not deleted when removing a mini-app from Layla - fixed a bug where LTM sometimes crashes when handling very very long messages - fixed a bug where images generated during chat were not being backed-up - fixed a bug where tags from TavernPNG were not being imported - fixed a bug where backup cannot copy references with spaces or special characters in their name - fixed a bug where automatically converting JPGs when saving as TavernPNG was giving an error - fixed a bug where personas were not being saved as part of your Inference Engine - fixed a bug where upscaling anime images caused a crash

- updated llama.cpp to support the latest iPhones

New features: - visualisation of temporal relationships in Layla's long-term memory A while ago, Layla introduced temporal memories: memories that evolve in time. This update brings a whole new way to visualise your character's memories! Your memories are now laid out in a 3D space; past memories are "deeper" into the z-axis. A slider at the bottom allows you to navigate through your memories in time, jumping from past to present. You will notice that memories naturally form "clusters" in time. These are memories that occur within a relatively short timeframe. Each cluster has it's own central node(s), with auxiliary memories branching out around them. These clusters are segmented via a time-dilated algorithm: recent memories are finely-segregated, while older memories are more clustered together. This mimics how human memories work; you often remember recent events clearly, while the longer the events are in the past, the more they combine into one big cluster. Important nodes in each cluster are linked together in time. These are usually "common entities" between temporal clusters. Intuitively, this should be important characters, topics, etc. The graph reflects this. The "anchors" in each cluster are usually the two main participants of the conversation (you and your character). These nodes glow brighter in the memory network, and you can see faint "temporal links" between them. These temporal links are annotated by the time elapsed between each cluster, which is visible in the memory graph. There are a lot more ways in which this marshalled data and relationships can be used during the recall process, and this is something I will continually improve on over time! Improvements: - significant improvements to the performance of chatting, especially for very long conversations spanning hundreds/thousands of messages - long-term memory will now recall earlier memories from the same conversation that has been cycled out through context shifting - fixed a few conditions in triggering new search agent - supports the new Ministral-3 family [THINK] tags Bug fixes: - fixed a bug where invalid filename in References mini-app causes Layla to crash - fixed a bug where characters list will sometimes show nothing when updating a filter - fixed a bug where character list resets to the first page after edit/saving a character - fixed a bug where regenerate messages leaves a large amount of empty space when the length of the regenerated message varies largely - fixed a bug where automatically searching for models in OpenAI compatible API endpoints was not working - fixed a bug where Layla does not stop the server from generating messages even after pressing "stop" - fixed a bug where voice chat does not speak aloud responses - fixed a bug where audio is only played through speakers and not AirPods

New features: - Layla supports Qwen3-VL family models: images are automatically down-sized before sending to the LLM to keep processing times reasonable on mobile, you can control how much down-sizing occurs in Advanced Settings - Layla supports animated PNGs and GIFs for your character cards Improvements: - added setting in Reference documents to control the threshold of how long a document needs to be before it is summarised instead of injected wholly into the context - prompt templates are now case insensitive - display character full name in chat instead of just the first name - automatically convert JPGs to PNGs when exporting your characters as a TavernPNG - you can share text with Layla to start a chat with that text Bug fixes: - fixed a bug where the document name was not detected correctly when adding a document from chat - fixed a bug where character sorting changes back to the default after exiting chat - fixed a bug where tool calls are included in long-term memory messages - fixed a bug where Layla crashes on importing a huge TavernPNG card

- fixed bug where option toggles are not showing

- improved stability in chatting - fixed a few bugs in edit/regen - fixed bugs related to the calendar agent