Improved web search result quality and allow user to see search links that were used by the llm Allow user to use embeddings on gpu for faster document process Improved suspension so app on iOS now do not get killed when backgrounded

bug fixes

- Conversation mode: hands-free voice chat with text-to-speech - Model setup simplified: configure defaults once at the app level - PDF support for RAG knowledge documents - Faster startup: models load in parallel, UI stays interactive - Manage Storage screen to view usage and clear data - CPU fallback warning when GPU models run on CPU - Better background behavior - Fixed Android crash, delegate toggle, and OAuth login issues - Security updates

• Load models and toggle delegated inference directly from Inference Details • Major sync reliability improvements — no more corrupted state after interrupted syncs • More stable delegated inference connections with better error handling • Faster chat startup: speech model now loads after the main LLM • Fixed duplicate onboarding, QR scan warnings, and app icon rendering • Pixel 10: GPU inference now allowed with a compatibility warning • SDK updated to v0.8.1

Improved chat experience with cleaner UI, better modal consistency, and updated microphone recording animation. Better delegated inference support, including clearer status indicators and improved compatibility with transcription and model selection. Fixed multiple connection and authentication issues (token refresh/keep-alive, reconnect behavior, and infinite loading edge cases). Resolved numerous bugs around project/thread naming, placeholders, permissions, and generation state feedback.

- Added one click integrations for Asana, Notion, AWS Docs, Chrome DevTools - Added built in web search - Fixed syncing with progress indicator - Added download speed and size - Fixed automatic reconnection to MCP servers at launch - Added approval step for cross device sync - Unified QR scanning for both sync and (opt in) delegated inference - Design revision + fixed icons - Reset data when revoking sync from base device and warning - Allow deleting cached models via sidebar - Add support for blind relays

This update introduces support for the Model Context Protocol (MCP). You can now connect your app to a world of external tools, services, and data, dramatically expanding what your local models can do. We've also included several bug fixes and stability improvements for a smoother experience.