On‑Device AI:Private Workspacev1.49.0

What changed

• PerformanceOptimized MOSS Transcribe + Diarize model to use less memory for long recordings.
• New featureAdded on-device OCR for scanned documents and PDFs without readable text.
• New featureIntroduced new audio models: Supertonic, VoxCPM2 Int8, VoxCPM2 BF16, and IndexTTS2.
• UXRedesigned Menu with Pinned and Available workspaces and adaptive Settings layout.
• UXSimplified Agent Workflow setup with Agent Profiles, AI Skills, and MCP servers.
• UXCentralized model management in the new Model Library for browsing and repairs.
• PerformanceImproved chat performance with smoother scrolling and early reading of long answers.
• PerformanceIncreased timeout for cloud requests and improved error handling for failed loads.
• PlatformAdded dedicated messaging app connection settings for Mac users.
• PerformanceFixed bugs related to saved chat loading and conversation mode listening behavior.

Original release notes

Version 1.49.0 is for the work that starts with a long recording, a scanned document, or an idea you want to hear out loud. It also makes a growing app much easier to find your way around.

• Recording a meeting, class, interview, or research session? MOSS Transcribe + Diarize is our recommended model for meeting notes. It separates speakers and now uses less memory, so long recordings can finish more reliably. Your original audio stays untouched.
• Add a scanned report, paper, receipt, contract, or class handout to AI Chat and ask questions about it. If a PDF has no readable text, the app asks before running OCR on your device. You can follow progress page by page or cancel without leaving a half-finished attachment.
• Supertonic, VoxCPM2 Int8, VoxCPM2 BF16, and IndexTTS2 are new in this release. Depending on the model and your device, you can make narration for a presentation, create a voiceover, listen back to your own writing, or use a saved reference voice.
• The Menu now keeps the workspaces you use most under Pinned, with everything else under Available. Settings are grouped by purpose and adapt to the screen, whether you are on an iPhone or working with a sidebar on a larger device.
• Agent Workflows are easier to set up. Start with an Agent Profile for a repeatable job, such as research or editing, then add the AI Skills, MCP servers, and tools it needs. On Mac, messaging app connections now have their own Settings area.
• The Model Library is now the main place to browse, download, choose, repair, or remove chat, transcription, speech, and embedding models. Speech downloads also show clearer progress, including quiet stretches during large transfers.
• Saved chats open correctly on the first try. Long answers scroll more smoothly, and you can begin reading before rich formatting is done.
• Cloud requests have more time to respond. If one fails, the app stops cleanly, keeps any partial reply, and leaves a successfully loaded model ready for another try.
• We also cleaned up the AI Chat model header and Voice Note controls. Conversation Mode returns to listening after it speaks, and deleting an active chat no longer lets old work bring it back.