Skip to content

1.9.0

Choose a tag to compare

View all tags
@andriydruk andriydruk released this 21 Aug 15:49
· 4 commits to main since this release
1.9.0
394ff98

What's New

Features

  • Documents — attach a file and ask questions about it: PDF, Word, EPUB, HTML, Markdown and plain text
  • On-device semantic search finds the relevant passages — files never leave the phone
  • A small embedding model downloads once on first use and is shared by every chat model
  • Tap a document chip to reopen the original file

Under the Hood

  • Embedding sessions in the inference engine, powering the on-device RAG pipeline
  • llama.cpp updated to b9978
  • Targets Android 16, dependencies updated
  • Store listing, README, landing page and screenshots updated for Documents, localized to all supported languages

Bug Fixes

  • Vulkan CLIP vision now runs only on an allowlist of GPUs (Adreno 6xx/710 denylisted, Mali-G710 allowed), falling back to CPU elsewhere
  • Tolerate duplicated vocab tokens in GGUFs instead of failing to load
  • End the turn instead of decoding an empty batch
  • Suppress the R8 error for the pdfbox JP2 codec
Assets 2
Loading