Changelog

Version 1.2 makes model status and context easier to understand, adds direct control over model reasoning, and keeps long chat libraries and streamed tool responses clearer and more reliable.

New

• Open the context indicator beside the composer to see used and remaining context, model limits and capabilities, latest input, output, and reasoning token counts, and relevant runtime settings.
• Choose whether compatible models use reasoning for new requests, independently from the existing Stream Reasoning setting that controls whether reasoning appears live.

Improved

• Streaming replies now show their current stage throughout the response, including prompt processing, model loading, thinking, tool calls, and response generation.
• The model picker now shows clear connecting and connection-failure states, offers a direct retry, and keeps its model list up to date while it is open.
• Chats now keep their model selection aligned with a loaded LM Studio model after refreshing the server, with clearer loading feedback when switching models.
• Recent chats and search results now continue loading as you scroll instead of stopping after the first 20 conversations.
• Markdown links, citations, tables, code controls, and text selection now adapt to the app’s chosen accent color and appearance.

Fixed

• Tool and integration failures now stop the active response, show the failure clearly, and preserve useful details instead of leaving the chat waiting indefinitely.
• Incomplete or empty terminal responses now mark unfinished tool calls correctly and avoid creating duplicate or unnamed placeholder tool rows.
• Assistant message timestamps now reflect when a response actually completed, failed, or was recovered after an interruption instead of when streaming began.

Version 1.1 makes LLMster easier to navigate, introduces smoother rich-text streaming and live reasoning controls, and improves reliability for demanding and background responses.

New

• Choose whether model reasoning appears live with the new Stream Reasoning setting. When disabled, LLMster shows a compact Thinking indicator while still saving the completed reasoning.
• Pin favorite models to the top of the model picker. Favorites are remembered across launches without changing the currently selected model.
• Choose between System, Light, and Dark appearance modes in addition to selecting a primary color.
• On iPhone and iPad, the chat library now separates Recent chats and Folders, adds search by chat title, model, or folder, and shows recent conversations from across all folders.

Improved

• A new streaming Markdown renderer displays responses and reasoning smoothly as they arrive, with animated text, native inline and block math, syntax-highlighted code blocks with copy controls, and tables that can be copied or shared.
• Transcript sharing now preserves the original Markdown, including headings, links, reasoning, tool calls, arguments, and output.
• Generated conversation titles can now use three to eight words, providing more context while remaining concise.
• Notification previews now remove Markdown formatting and limit unnecessary processing for unusually long responses.
• Local Network settings now link directly to setup instructions, and several settings controls and the Chat Settings icon have been made clearer.

Fixed

• Reworked response transport and presentation to address UI lockups and watchdog crashes seen with fast, reasoning-heavy, and QAT model output.
• Background responses on iPhone and iPad now use their fallback background time only after the app enters the background and clean up correctly when completed, cancelled, expired, or brought back to the foreground.
• Interrupted responses now preserve partial output, mark unfinished tool calls correctly, and recover cleanly after the app is relaunched.
• Improved the handling of very long streamed responses and incomplete-response notices without repeatedly processing the entire message.
• LM Studio authentication key changes are now saved securely and applied immediately without leaving stale model information in the chat.
• Fixed keyboard focus sometimes remaining active after first-run server setup was closed.