v1.1.1: Local Mode
Immutable
release. Only release title and notes can be modified.
v1.1.1: Local Mode
Run the entire conversation on your own hardware. Three independent toggles in Settings route the LLM, voice synthesis, and transcription to your own machine. With all three on and a local LLM, no conversation, microphone audio, or TTS audio leaves your machine.
Highlights
- Local Mode panel with per-service toggles for LLM, TTS, and STT. Mix and match. Most self-hosters keep the LLM on OpenRouter and run just voice and transcription locally.
- Docker sibling containers for English voice (Kokoro CPU image), German voice (Qwen3-TTS CUDA, behind the
nvidiacompose profile), and transcription (Whisper large-v3-turbo, ~5x faster than large-v3 on CPU). - Apple Silicon native MLX paths. Both Qwen3-TTS (German voice) and Whisper STT run on Metal via dedicated FastAPI wrappers. About 20x faster than the CPU docker path for short utterances. Setup is one command per service.
- Marketing pages migrated to Astro for sub-second first paint on figure and theme pages.
- Legal pages (Impressum, Datenschutz, Cookie-Richtlinie, Nutzungsbedingungen) shipped as Astro with correct canonical URLs,
og:locale=de_DE, and per-URLlastmodfromgit login the sitemap. - Voice technology credits on
/aboutnaming the upstream projects we depend on: Qwen3-TTS, F5-TTS by SWivid et al., the hvoss-techfak German fine-tuning, Kokoro by hexgrad, and faster-whisper served via Speaches.
Companion repo
chipmates/f5-server (publishes within 48 hours of this release) packages our production F5-TTS deployment as a standalone OpenAI-compatible service. MIT-licensed code, bring-your-own-checkpoint.
Setup
Three steps in docs/SELF-HOSTING.md: docker compose up -d for the app, two toggles for local voice and transcription, and one optional step for a local LLM with our recommended 27B-class GGUF.