Atmee v2

Atmee v2 is Atmanity's real-time video avatar foundation model (sites atmanity.io / atmee.ai), launched on 2026-11-04 by Chen-Ping Yu's team (Phiar was acquired by Google).

Versus shoulder-up v1, v2 renders a full frame — face, body, micro-expressions, emotion and environment — in real time on a single GPU. Companion model Causeur drives behavior and personality. Build an avatar from one portrait plus a 10–60s voice clip; APIs support full-stack, avatar-only and voice-only modes.

What problem it solves

Support, tutoring and game NPCs want face-to-face AI, but general video models are slow/expensive and classic avatars look stiff. Atmee v2 targets cheap real-time conversational video agents.

Highlights

• Set engine=v2 on create; default is v1.
• Example pricing: full-stack v2 ~2 credits/min, avatar-only v2 ~1 credit/min (see docs).
• Takes: capture multiple previews and pick which goes live.

Limits

• v2 needs Growth+ or an invite.
• Real-person likeness/voice requires permission.
• Early access — validate reliability and compliance yourself.

Who it's for

• Product teams building live support/tutor/companion video agents;
• Developers with a voice agent who only need synced face rendering.

Getting started

1. Visit atmee.ai and the create-avatar docs.
2. Get v2 access (Growth or invite) and an API key with avatars:write.
3. Create via JSON manifest or zip with portrait + voice and engine=v2.
4. Start full-stack sessions with POST /v1/sessions, or use the LiveKit plugin for avatar-only.