SmolChat-Android

• The app allows benchmarking the model used in a chat. Go to Settings > Benchmark Model in any chat and click 'Start Benchmarking'. The benchmark results are pp (tokens/sec for prompt processing) and tg (tokens/sec for token generation).

• Add support for Portuguese language (alongside existing support for English and simplified Chinese).
• Clear model context/memory when clearing the chat i.e. deleting all messages from the current chat.
• Avoid model reloading when screen is rotated.

UI Improvements

• Earlier, the user had to click on the folder name to expand it. Clicking on the chevron icon now also expands the fol...

• Fixed a bug where the app's memory usage kept increasing after switching models i.e. the memory acquired by the previous model was not 'released' when selecting a different model
• Sync with upstream llama.cpp
• Align default inference parameters with those found in llama executable

UI Improvements

• Chat message actions like share/copy/edit are now available in a dialog which appe...

• Sync with llama.cpp upstream
• The app now uses a new set of icons for a more aesthetic/refreshed look
• Fixed a bug where the app's memory usage kept increasing after switching models i.e. the memory acquired by the previous model was not 'released' when selecting a different model

• fix bug in ChatActivity causing a NullPointerException when the app is launched (this is the bug causing most crashes according to Google Play)
• make 'Download Models' screen scrollable to make sure it works correctly on small screen devices
• add support for 16 KB page sizes
• improve the SmolLM API and document its methods