2.2.1 — critical download fix
🛠️ Fixed: alternative models downloaded corrupted. The alt-quant models (Gemma Q5/Q6/Q8, Higgs, Roformer, Parakeet) moved to Hugging Face's new Xet storage, which drops connections under parallel download — the files came down with holes and llama-server failed to load them (invalid magic characters '????'). Now they download correctly:
- range retries with backoff, lower connection concurrency for Xet;
- download to a
.partfile, validated (exact size +GGUFmagic) and only then renamed — a corrupt/interrupted download can no longer pass as "installed"; - fixed a Parakeet manifest typo (
encoder-model.onnx_data→.onnx.data).
Verified end-to-end through the app: all alternative models download valid, and the Gemma Q8 quant loads and translates in the real dubbing pipeline.
🎬 Transcript mode — karaoke play-along: a play button, video, and highlighting of the current line/word as it plays.
🌍 Processing status now follows the interface language (was hardcoded Russian).
🔄 Auto-update: 2.1.0 / 2.2.0 will offer to update.