I've been working on this after hours for a few months continuously improving it. Now at a point where the chatbot is close to real-time (thanks to SSE streaming) and also interruptible while preserving context of what was last said. 100% local and powered by Qwen3.5-397B (Unsloth's UD-Q3_K_XL), Whisper-small STT, and Orpheus Q4_K_XL TTS with a custom SNAC decoder on ONNX. VRAM usage holds at 21.3 GB or less leaving decent headroom for compute graphs on a 24 GB GPU. System RAM MoE experts for ...