Meta opens Muse realtime AI video calls to all US users at 870ms latency
rohanpaul_ai · x · 2026-11-06
Meta opened live video calls with its Muse AI agent to every US user: callers see an animated avatar (restyleable as person/animal/character) that reacts while listening.
Key technical details:
- Muse Realtime Voice emits speech tokens carrying words and delivery; Muse Realtime Avatar turns the same stream into lip-synced video
- A student model reproduces in 2 passes per video chunk what a teacher model needed 120 passes for
- Streams 448x768 portrait video at 25fps, starts replying 870ms after user stops talking
- With 4-bit quantization, caching and dynamic batching, one Nvidia GB200 runs 12 concurrent video streams
In company-run 2-3 minute call tests, raters preferred Muse over Runway Characters 78% to 22% and HeyGen LiveAvatar 88% to 12%. At 12 streams per GB200, 1M simultaneous callers would need 83K GB200s for video alone — serving cost could cap the avatar's reach.
Related event: Meta Rolls Out Muse Voice Calls and Realtime Avatars to All US Users(10 posts)→
More from Multimodal
- World Labs world model demo shows actual moving people in generated scenes — andrew_n_carr · 2026-11-06
- Grimes stars in first AI film to hit theatres, gunning for Best Animated Feature Oscar — TomLikesRobots · 2026-11-06
- AI-generated video now looks like full TV show episodes — Additional_Judge_337 · 2026-11-06
- AI-generated Willem Dafoe Green Goblin shots show how far image models have come — egeberkina · 2026-11-06
- How Grimes' cameo in theater-bound AI film 'Gods Don't Give Gifts' came together — DavidmComfort · 2026-11-06
- Is LTX 2.5 broken in ComfyUI? Users report eerie, degraded outputs — brocolongo · 2026-11-06