LastChat

• Added the ability to use multiple api keys on one llm provider
• Added indications to the local model import process
• Fixed the Keyless search provider
• Fixes for stuff that could have triggered crashes
• Removed GitHub Models from the LastChat catalog

Make sure to backup before updating!!

After a long wait, here's LastChat 1.4.6! This version has a lot of changes, so some of them may not be fully polished yet. Just keep that in mind.

Additions:

• Support for local LLM and embedding models through LiteRT-LM
• Support for various local speech-to-text models through sherpa-onnx
• Deeper Android integration by allowing you to choose LastChat as the system assistant
• UI to quickly check the currently selected model's context size and context usage
• A smarter context management system that will probably replace the old per-character context management settings in the future
• OAuth MCP server support
• Added token capacities and custom reasoning modes to the LastChat catalog
• Queued / gapless TTS playback for longer replies

Improvements:

• Character card import/parsing is more robust
• Binary / huge files aren’t shoved into the prompt anymore (you’ll get pointed toward a workspace instead)
• General improvements to ui consistency and other stuff
• A lot of other things I forgot

• Workspaces have been added to the tablet mode settings
• The update process has been reworked to be more seamless and also work in tablet mode
• Decreased possibility of stutters when scrolling in a chat
• Added fallback RSS mode to the Bing search provider
• Added workspace support for ARMv7 devices
• Fixed an issue where provider presets would autonomously get added to the user's provider list
• Some small fixes to the memory system

The new update popup is inspired by the Nuvio update process

Make sure to back up before updating! This release changes the local workspace architecture, contains a database schema migration, and replaces sandbox file handlers, it's recommended that you backup your data before updating.

Visible changes:

• The Chaquopy Python sandbox got fully replaced with Workspaces/PRoot-based Linux environments
• Added support for OpenAI-compatible speech-to-text models
• Added some new providers to the catalog
• Lorebooks can now be toggled per-chat too
• The activity timeline got redesigned to better match the rest of the UI
• The activity pill in the webUI got refined to look more like the in-app counterpart
• Provider and model icons in the webUI should now better match their in-app counterparts
• Added a fullscreen mode to the message input field when there's a lot of text
• The main settings screen got simplified
• Other small refinements

Hidden changes:

• Replaced standard Java I/O streams with platform-agnostic okio buffers, sources, and sinks across file compressor routines, database backups, media encoders, web uploads, and attachments
• Removed Java concurrency utils (such as ConcurrentHashMap and thread locks) in favor of portable Kotlin Atomic references and Mutex locks for throttling, provider balances, and file storage mappings
• Introduced PlatformHttpClient and platform JWT signers to facilitate portable network handling outside of Android-specific library contexts
• Replaced java.time.Duration references with Kotlin's native Duration class for timeout parameters and scheduled task definitions
• Extracted system logging (PlatformLog), haptics (PlatformHaptics), syntax highlighting labels, and media encoding logic to remove direct dependencies on the JVM
• Added boundaries on database fetches and DAO actions to prevent freezes and memory exhaustion
• Capped the number of assistant memory entities retrieved during RAG fetches (limited to 50 for empty/simple queries, and coerced at a maximum database limit of 100 to 200)
• Swapped in-memory filtering with database-level constraints (getMemoryEntitiesOfAssistantLimited directly in Room) to avoid loading massive history payloads into JVM heap memory. This should prevent app stuttering and freezes during long conversations
• Simplified the internal Android cache stores (PerKeyFileCacheStore and SingleFileCacheStore) by replacing heavy Java ReentrantLock structures with lightweight Kotlin/JVM monitor locks
• Optimized prompt formatting to maximize caching

Other stuff I'd like to mention:

• The APK size is less than half of what it was before because the Python sandbox isn't bundled with the app anymore

Found an issue after release: The workspace feature and update banner aren't currently supported in tablet mode. I will release an update with these fixes as soon as possible!