MynahOpen-source digital humans, ready to deploy.

One portrait, one command — a digital human that talks back in realtime, reacts with gestures, and ships as a link. Engine, console and delivery path, all open source.

Quickstart (15 min)

Architecture

GitHub

Realtime, not rendered video

The cloud brain answers within 0.2 s of your last word and lip-synced video starts under a second; interrupt it and it stops. A semantic end-of-turn model decides when you are actually done. WebRTC straight to the browser.

One portrait, a half-body presenter

FlashHead renders the whole head and pastes it onto a 720×1280 head-to-waist canvas; idle breathing and gesture clips (shy, laugh, shake, wink) are baked from the same canvas. Luna and Nova ship as presets.

Instinct, not just answers

An instinct layer picks an expression or gesture while the brain is still composing, and plays it the moment the reply ends. Backed by the Jev cognition model, or a zero-dependency rules provider when offline.

A console, not a script

Avatars, voices, knowledge bases, channel publishing, access control, concurrency quotas, an engine pool with a VRAM budget, one-click avatar swap per channel.

Plugs into your stack

Brain options — local LLM, Qwen Realtime, or any Pipecat pipeline; avatar, TTS and ASR behind gRPC and OpenAI-compatible contracts; Python SDK and a DeepSeek Harness plugin included.

Yours to ship

Apache-2.0, no feature wall, no watermark requirement, consent-gated avatar training. One RTX 4090 runs MuseTalk and FlashHead side by side; every component can also be a remote endpoint.