One layer. Four steps.
The iOS app sends speech → the server runs STT, RAG, SLM and TTS in sequence and streams the answer back as audio, sentence by sentence. Every model runs on Alibaba Cloud Qwen; the API key stays on the server.
🎤 iOS→API→STT→RAG→
SLM→TTS→🔊 ~2 s
Models
Gateway: ali.gerege.mn/v1 (OpenAI-compatible). Measured on 2026-11-03.
| Role | Model | Speed |
|---|---|---|
| STT | qwen3.5-omni-flash | ~1.7 s |
| RAG | text-embedding-v4 (1024d) | ~0.4 s |
| SLM | qwen3.6-35b-a3b · MoE, 3B active | first token ~0.5 s |
| TTS | qwen3.5-omni-flash · Voicebox (clone) | first audio ~2.1 s |