🚀 New Features
- vLLM managed GPU engine — run and manage vLLM as a second managed GPU engine alongside existing engines.
- Linux ARM64 builds — added support for building Linux ARM64 AppImages and related ARM64 build tooling.
- atomic-prism provider support — added the atomic-prism extension and provider integration so Prism-only models are recognized and handled.
🔧 Improvements & Fixes
- Documents are cited by readable labels (e.g. "[FINDINGS.md §13]") and citations show a clickable passage card with file, passage text, Open document and Show in folder, or removed state (ATO-551).
- Retrieval now accepts up to 5 queries (one per fact), raises default top_k from 3→5, and merges query embeddings in linear search so multi-fact questions return the correct passages (ATO-552).
- Finished replies no longer block the page composer after remounts — per-thread session callbacks and request/tool aborts were fixed so Enter/Stop and composer state behave correctly (ATO-538).
- If a chat model load is refused while an image/video job is running, the app now shows "The GPU is busy" with the job name and guidance instead of a generic "Failed to load the model" message (core 0.11.4, ATO-549).
- Stop now closes the local streaming connection in Rust and frees the engine slot so subsequent messages are not blocked by abandoned streams (ATO-550).
- Video page wording and API copy fixed — API endpoint copy, requirements, and runtime controls now refer to "video" where appropriate (ATO-553).
- Hub and provider pages show more embedding model details — dims, Matryoshka lengths, configured context and example prefixes; llama.cpp embedding GGUFs are listed under "From your llama.cpp models", and provider state now shows Loaded/Idle/Starting/Failed/Stopped (ATO-544/546/547/548).
- Decision model starts that timed out or failed now offer Retry and View logs instead of incorrect advice, and failed engine-update checks are no longer reported as "up to date" (ATO-542/543).
- Desktop internal links are kept in the app router so same-origin internal links open inside the app instead of being sent to the system browser (UI fix).
- 10+ additional stability and UI fixes and test quality improvements.
🙏 Contributors
Thanks to @Ooooze, @Vect0rM for their contributions to this release!