The example builds a video avatar using BitHuman’s open-weights Expression Video Avatar and Qwen3-TTS, then connects it to a LiveKit room.
The setup uses Modal’s Server endpoint to launch the avatar container. Modal says its Sticky Sessions primitive keeps a client associated with the same container after WebSocket reconnections; the LiveKit agent starts the session, calls the /launch endpoint with its token, and ends the session when the call finishes.
Modal also says GPU snapshots hide cold starts that would otherwise require loading gigabytes of model weights. The post links to full code and describes the BitHuman integration as a self-hosted GPU container.
The source does not provide latency, throughput, reliability, pricing, or production benchmark results for this configuration. It also does not announce a new LiveKit or Modal product release.