## v1.1.035 — 2026-05-29
- fix: **HQ Voice (voice-cloning) is no longer wrongly blocked from starting when GPU memory looks tight.** Selecting HQ Voice could be refused with a "not enough GPU memory" notice — sometimes even when several GB were actually free, and inconsistently from one launch to the next. That pre-emptive guess has been removed: HQ Voice now starts whenever you select it, and if there genuinely isn't room when it goes to speak it just reports that, rather than refusing up front based on a wrong estimate. The voice engine's memory is still accounted for when loading the local AI model, so model loading stays protected.
- fix: **The "free up GPU memory and retry" recovery now actually works when a local AI model fails to load.** If loading the local AI model ran out of GPU memory while HQ Voice was running, the recommended one-click action that was meant to release the voice memory (~2 GB) did nothing — a leftover from an earlier change to how the voice engine runs — so the retry kept failing the same way. It now properly frees the voice memory, falling back to Fast Voice, before retrying the model load, and the Voice Output settings update to show the switch (so testing the voice no longer re-triggers the out-of-memory error).
- improvement: **Cloud API provider order.** The pinned top row now shows Google Gemini, Groq, OpenAI, and Custom (OpenAI-compatible). Every other provider stays in the second section in alphabetical order, as before.
