Meta Rolls Out Real-Time Video Chat for Muse AI Avatar Across U.S.

Meta has begun a nationwide U.S. rollout of Muse Voice, a feature that lets consumers hold live video calls with an animated AI avatar. The experience is driven by two newly disclosed models — Muse Realtime Avatar and Muse Realtime Voice — which synchronize facial expression, lip movement, and prosody in a single low-latency pipeline. Unlike pure voice assistants, the system presents a persistent, customizable character (nicknamed Jolly) that users can reshape into human, animal, or fantasy forms.

The company frames the anthropomorphic layer as a retention lever: a visual counterpart that mirrors conversational tone through subtle gestures is intended to deepen user attachment and increase daily active sessions. Chief AI officer Alexandr Wang demonstrated the capability in a public demo where his dog-styled avatar responded to a joke about meme policy with a contextual suggestion, illustrating the model’s ability to blend humor, safety awareness, and task execution in one turn.

Why it matters for GPU / AI infrastructure

Real-time avatar inference demands sustained sub-100 ms frame generation at 30–60 fps, combined with streaming voice synthesis — a workload profile that stresses GPU memory bandwidth and tensor-core utilization differently than batch LLM serving. Meta’s disclosure that the stack runs on mobile, web, Mac, and upcoming AI Glasses plus a "Muse Charm" wearable implies a heterogeneous deployment target spanning cloud GPUs, edge silicon, and on-device NPUs. Infrastructure teams should watch for reference architectures that optimize model partitioning across this continuum, especially as the wearable form factor introduces strict thermal and power envelopes.