Meta is rolling out Muse Voice, a live avatar chat for its AI agent, available to all U.S. users. The feature, which lets users video chat with a customizable animated avatar named Jolly, operates on mobile, web, and Mac, with plans to extend to Meta’s AI Glasses and the upcoming Muse Charm wearable. TechCrunch reported the launch, noting the timing precedes Meta’s year-end release of a Tamagotchi-like device for the avatar.
Behind Jolly’s friendly interface lies a complex system. The avatar is powered by Muse Realtime Avatar and Muse Realtime Voice, technologies Meta first demonstrated at its Connect conference in September. The system processes a single stream of data to generate synchronized speech and facial expressions, using a diffusion transformer model to animate the character. This architecture allows for a claimed latency of approximately 870 milliseconds from the end of a user’s turn to the first synchronized response, according to explainx.ai, which analyzed Meta’s technical research paper.
Muse Realtime Avatar: Synchronized Audio-Visual Generation
The core of Muse Voice is its unified audio-visual pipeline. Unlike systems that treat voice and video as separate streams, Meta’s approach uses a single stream of speech tokens containing both content and delivery instructions. This stream feeds into two parallel decoders: one producing the audio and the other generating the avatar’s synchronized visual output. The model conditions its output on reference media—be it a photographic portrait, an illustration, or even an animal—and a rolling window of recent video data to maintain context. This is achieved through a diffusion transformer model that animates the character in short, causal chunks, with each completed segment providing motion context for the next.
To make this process computationally feasible, Meta employs a sophisticated distillation technique. A bidirectional teacher utilizes 40 diffusion steps combined with three-way classifier-free guidance, equating to 120 model evaluations per chunk. This process is then condensed by Meta into an unguided two-step student, representing a 60x reduction in evaluations, through the application of self-forcing and distribution matching distillation. This is critical for achieving real-time performance on consumer devices.
System Generates Eight Frames Fast
Meta’s public benchmark for the system is an 870-millisecond latency from the end of a user’s utterance to the first byte of the avatar’s synchronized response. explainx.ai notes this figure is derived from Meta’s own research paper, though independent verification of this claim is currently unavailable. The company states that running on a single NVIDIA GB200 accelerator, the system can generate eight frames of video (320ms of playback) in just 20 milliseconds. This performance is key to making the conversation feel fluid and responsive rather than disjointed. The system streams video at 448×768 resolution at 25 frames per second, a resolution chosen to balance visual detail with the demands of real-time generation.
What This Means for Developers and Third Parties
Currently, the Muse Voice system is a closed consumer product. For now, the focus is on consumer engagement, with Meta betting that an anthropomorphized, customizable avatar will create a stronger emotional connection than a disembodied voice.
The rollout’s timing is strategic. The Tamagotchi-like Muse Charm device, slated for before year-end, will rely entirely on the user’s established connection with the digital avatar to drive adoption.
While the underlying large language model is not new—the core models were first shown on September 23, 2026—the innovation lies in the real-time rendering and synchronization pipeline that brings the avatar to life.
From Digital Companion to Wearable Device
The current software rollout is the first step, with the Muse Charm wearable representing the next. This small device, described as “Tamagotchi-like,” will feature a screen displaying the user’s customized Muse avatar, allowing for voice conversations and task delegation on the go. TechCrunch confirms the device is planned for a before year-end launch.
🚨MUSE FEATURE ALERT 🚨 muse voice is out for everyone in the US! you can now phone call and video chat with your muse whenever you need. i personally have a sticky situation today, and asked muse for some advice. sage and calm as always ❤️ pic.twitter.com/zTtLa6R1EJ
The customizability of the avatar is central to this strategy. Users can morph Jolly from a default cute character into a photorealistic human, an animal, or any other image they provide. This personalization is designed to foster a sense of ownership and attachment, making the AI feel less like a utility and more like a digital pet or a personal assistant with a unique identity. As Meta’s Chief AI Officer, Alexandr Wang, demonstrated in his own demo—where his Muse appeared as a fluffy white dog offering meme advice—the system is built to handle casual, conversational interaction with a degree of personality.
Muse Voice Makes Interaction Visual
Muse Voice makes AI interaction more visual and engaging.
