What's new in v0.27.1
A new local translation model for low-resource languages, plus a TTS fix for drifted system clocks.
🎉 New — Hunyuan MT 1.5 1.8B local translation (36 languages, WebGPU)
Tencent's translation-specialized HY-MT1.5-1.8B (WMT25 championship lineage) is now available as a local translation option. A single model covers 36 languages including low-resource targets that previous local options couldn't handle: Khmer, Burmese, Tibetan, Mongolian, Uyghur, Kazakh, plus the usual Chinese / English / Japanese / Korean / Romance / Arabic / Hindi etc.
It runs entirely in the browser via WebGPU — no API key, no server round-trip.
How to use: Settings → Translation → Local Inference → Model Management → download "Hunyuan MT 1.5 1.8B".
- q4 (~1.34 GB) — works on any WebGPU device
- q4f16 (~1.17 GB, faster) — requires
shader-f16(most newer GPUs auto-detected)
Closes #233.
🐛 Fixes
- Edge TTS no longer fails when your system clock is drifted. Microsoft Edge TTS requires a request signature derived from the current time; if your machine's clock was even a few minutes off, auth would silently fail and Edge TTS voices would never play. Sokuji now reads the server's
Dateheader on the voice-list probe and compensates for the local-clock offset on every subsequent request, so Edge TTS works regardless of clock skew.
Full Changelog: v0.27.0...v0.27.1