HeyGen open-sources a real-time avatar stack under an MIT licence

HeyGen published its September 2026 release notes on 1 October, covering three items.

On 10 September the company released an open-source real-time avatar stack that integrates GPT-Live-1 with HeyGen LiveAvatar and HyperFrames overlays. It ships with three working demos — a language tutor, a poker coach and a customer support agent — and both repositories are MIT licensed and available on GitHub.

On 18 September HeyGen launched Code2Video on Kaggle, a benchmark measuring how AI models convert creative briefs into motion-graphics code. It contains 168 briefs, each described as a beat of a real launch video, and judges output across five axes: engagement, prompt-intent, composition, temporal and craft. The company says a purpose-built judge model reaches 82% accuracy against human raters.

Also in September, HeyGen made a Professional Voice Clone API available. It requires a minimum of 20 minutes of recordings from a single speaker, is priced at 0.6 API credits per generated minute, and supports retraining without losing the existing voice identity; each slot includes five trainings per billing period.

Why it matters for digital humans: An MIT licence on a working real-time avatar stack is unusual from a commercial vendor at HeyGen's scale, and it lowers the cost of evaluating real-time avatars from a procurement exercise to an afternoon with a GitHub clone — worth noting that the open repositories still sit on top of HeyGen's hosted LiveAvatar and model services, so this is open tooling around a closed core, not a self-hostable pipeline. The voice clone spec is the more immediately useful number: 20 minutes of single-speaker audio is a real consent and production burden if you are cloning a named executive or instructor, and the retraining-without-identity-loss behaviour is what determines whether you can refresh a voice later without re-recording the talent. The 82% judge-model accuracy figure is HeyGen's own and relates to its own benchmark.