Local multimodal inference and reliable PDF processing
Local models can now understand image input, with more reliable PDF extraction dependency loading in the desktop app.
## Local multimodal inference
- Luxi Little and Luxi Pro can now install their pinned, checksum-verified vision projection files and understand image input while using local models.
- Model download, resume, verification, and legacy-file migration are more robust; a verified main model is not downloaded again when only its vision projection is missing.
- Chat and Agent continue to share the same local inference gateway, and image processing stays on the device without automatic cloud fallback.
## PDF processing reliability
- The PDF text extraction dependency is now explicitly preloaded and verified during desktop packaging, reducing missing-dependency failures after installation.
- The PDF extraction skill has clearer input validation, error handling, and output constraints.
## Installation
This build is ad-hoc signed, is not signed with an Apple Developer ID or notarized by Apple, and does not include an updater artifact. Download and install it manually from the Luxi website. On first launch, macOS may ask you to confirm Open Anyway in System Settings β Privacy & Security. You do not need to disable Gatekeeper.