Meta gives Muse a live avatar, with subsecond video responses

Why it matters

Meta is testing whether its Muse assistant can make generated video part of ordinary conversation. The launch's key evidence is engineering-focused: Meta reports about 870 ms to the first synchronized response and 12 video sessions per GB200, while its preference comparison remains company-run.

Alexandr Wang (@alexandr_wang) introduced Muse Realtime Avatar on September 24th, adding a live, animated face and body to Meta's Muse assistant. The model synchronizes generated video with Muse Realtime Voice, and Meta says the system returns the first byte of a synchronized response about 870 milliseconds after a user finishes speaking. Wang, Scale AI's founder who joined Meta to work on its AI efforts, is positioning the avatar as part of a broader push to make Muse a conversational agent people can see as well as hear.

https://x.com/alexandr_wang/status/2102919552254484765

In his X thread, Wang said Muse Realtime Avatar animates a character while it speaks and can sustain a conversation. He also said Meta tested it against Runway Characters and HeyGen LiveAvatar, using each product's live-call experience with the same character, and that participants preferred Muse. The Meta Research write-up, published September 23rd, describes two- to three-minute conversations with matched avatar identities. Meta says raters preferred Muse overall and across the evaluated dimensions, though its comparison of mannerisms with Runway was statistically indistinguishable from parity. These are results from Meta's own evaluation, not an independent benchmark.

The system's technical bet is to generate speech and video from a shared stream. Meta says Muse Realtime Voice produces speech tokens that carry both spoken content and delivery; the avatar model uses those same tokens to generate facial expression and movement, keeping lip motion and voice aligned. Its video model is an audio-driven Diffusion Transformer that generates short chunks and carries forward recent video context to maintain an avatar's appearance and mannerisms during longer conversations.

That approach targets a familiar weakness of generated video: a character can drift visually as a clip continues. Meta says it trains a causal model on its own generated context, then distills a teacher model that requires 120 evaluations per chunk into a two-step student. The company describes that as a 60-fold reduction in model evaluations. The operational result it reports is 448-by-768 portrait video at 25 frames per second, with about 870 milliseconds between the end of a user's turn and the first byte of the synchronized voice-and-video response.

Meta also says its serving optimizations let one Nvidia GB200 handle 12 concurrent video-generation sessions, eight times the capacity of its two-step BF16 baseline. That figure is a hardware-specific serving measure, not a published cost per conversation or a measure of capacity across Meta's full infrastructure. Meta says it worked with Nvidia on model optimizations and rebuilt its real-time inference stack around caching, dynamic batching, four-bit quantization-aware training, fused kernels and CUDA Graph capture.

The launch fits Wang's move from building AI infrastructure to overseeing models and consumer-facing products. Before joining Meta, Wang founded Scale AI, after work as a technical lead at Quora and roles at Hudson River Trading and Addepar; Scale's own profile says he studied artificial intelligence at MIT and left before graduating. Meta announced in June 2025 that Wang would join its AI effort as part of a deal in which Meta made a major investment in Scale. Muse Realtime Avatar puts that leadership closer to the user-facing product: a model feature whose appeal depends on reliable, low-latency interaction, not just on the quality of a generated clip.

The consumer-product boundary remains important. Meta's post says its examples demonstrate model capabilities and do not all represent avatars available in the Muse app; it also says Muse is for users aged 18 and older. The company describes the model as able to animate any reference image, including portraits, full-body illustrations, animals and everyday objects, but that breadth should not be confused with a promise that every example can be used in the app today.

Wang also says generated video is watermarked as AI without adding latency. Meta's research post specifies that the system uses Video Seal to embed a durable, invisible watermark. That makes provenance part of the real-time product design, although a watermark addresses traceability rather than questions of consent or how viewers interpret a lifelike avatar. For Meta, the launch combines a capability demonstration with an infrastructure test: it needs the assistant to respond quickly, keep a character visually consistent, and serve multiple sessions at once. The reported latency and throughput offer evidence on those engineering targets; Meta's own preference test supplies the product case.