Atmee v2 Explained: Real-Time Avatars, Claims, and Access

A face for the interface after text and voice

The interface after text and voice starts today.

Chen-Ping Yu says the idea followed Google’s acquisition of his AR startup. Text and voice already work, he argues, the way a command line worked before a graphical desktop, while nothing feels more intuitive than talking face to face. After two years of research, Atmanity launched Atmee v2 on November 4, 2026 as a real-time avatar model for live video agents that are meant to feel life-like.

At its core, Atmee v2 is described as a video foundation model that can render broadly, while the product is built to excel at characters. That is the job Yu says people mostly give video models such as Seedance 2.5 and Wan 3.0. The pitch is practical agents you can see: tutors, sales coaches, support agents, companions, game characters, and even whole living worlds.

What one image and ten seconds of audio are supposed to unlock

Any character from 1 image + 10s of audio.

The launch description is specific about the starter kit. Any character is said to come from a single image plus about 10 seconds of audio. From there, Atmee v2 is meant to drive contextual expressions and gestures, and to interact with the environment, rather than playing back a static talking head.

Three ways of using that character are named: live video calls, voice only, and video generation. Live calls are the headline, because the model is framed as real-time. Voice-only keeps the same agent when a camera is unnecessary. Video generation is the offline path when you want rendered footage instead of a call. Architecture details beyond “video foundation model,” including parameter count, training data, latency numbers, and resolution, were not part of the preview.

The speed and price claims, and who they are aimed at

Up to 50x faster. 10-100x cheaper.

Set next to character-heavy video models, Atmanity’s comparison is blunt: Atmee v2 is up to 50 times faster and 10 to 100 times cheaper. Those figures are the creator’s claims from the launch notes. Independent benchmarks, the exact models in the cost table, hardware, resolution, and per-minute pricing were not published with the preview, so treat the multiples as a positioning statement until an API quote or a public eval exists.

The intended users are teams building people you talk to, not one-off clips. A tutor who reacts while you stumble through a problem, a sales coach who runs the same role-play on repeat, a support agent with a consistent face, a companion, or a game character that can sit inside a larger world all fit the stated brief. Cheaper, faster character video would matter most where calls run for minutes and many users at once, which is also where unverified cost claims need a real invoice before you plan a budget.

What is actually available right now

No public build, download, or official page has been published. So far the creator has shown Atmee v2 in a short preview and the launch announcement. Editor notes on the release describe the API as invite-only. Open weights, a self-serve signup, app stores, and a documented price list are not in the materials.

Until an invite lands, you cannot time a call, check whether one photo and 10 seconds of voice really hold up, or see how environment interaction behaves outside the preview. Open questions sit next to the ambition: how “real-time” feels on a normal connection, how far contextual gestures go, how the model handles identity from a single image, and whether the 50x and 10–100x comparisons survive outside Atmanity’s own framing.