v0.8.4a

Quick patch:

• Updated cui-llama.rn DeepSeek R1 model support.
• Added default DeepSeek Instruct format.

v0.8.4

A small update with minor fixes and changes. Most changes are on the backend for maintainability purposes and preparation for future UI and localization reworks.

Changes:

• Updated cui-llama.rn - This should properly add support for Falcon models and fix a few issues with performance.
• Chats now export to .json instead of .txt.
• Sampler presets have been moved to a Zustand state rather than the ancient json file system. Sampler presets should automatically be migrated to the new system. This doesn't change much on the user-side, but may result in slightly faster transition time to the Sampler menu.

Fixes:

• swipe buttons being off-screen for smaller devices
• file sizes for models being cut off at 2GB - models will need to be reimported to fix this, but this bug is purely visual and does not impact performance.
• Flashing 'No Characters Found', index page and 'No Models Found' pages no longer appear.
• Slider values not clamping properly, leading to values such as 10.0000001 appearing.
• Context Length for Chat Completions being calculated completely wrong.
• Reimplemented missing 'Bypass Context Length'.

Dev:

• prepared i18n packages and preparing for future UI and localization reworks
• refactored entire TTS state to be seperate from component
• developer builds now use a different app id to allow parallel installs with release version