Meituan releases LongCat-Video-Avatar 1.5, its audio-driven video avatars

Meituan open-sources LongCat-Video-Avatar 1.5, an audio-driven avatar framework using a 13.6B parameter model to rival HeyGen and Kling Avatar 2.0.

Meituan open-sources LongCat-Video-Avatar 1.5, a framework for generating animated video avatars from an audio signal. The system relies on the LongCat-Video foundation model, with 13.6 billion parameters, designed to produce long videos, on the order of several minutes, without color drift or loss of quality.

The stated objective is to move beyond few-second demonstrations to achieve stable and consistent sequences over time, with meticulous lip synchronization, including in challenging situations such as multi-character scenes, object manipulation, singing, or animated styles. Key improvements in this version include the audio encoder being upgraded to Whisper-Large-v3, which refines mouth movements; character identity gaining stability in long videos; and inference being achieved in just eight steps through distillation, with an INT8 option to reduce memory requirements. The model accepts one or more simultaneous audio tracks.

In terms of results, the team reports a human evaluation conducted on 508 image-audio pairs by 770 testers and a dozen experts, across six scenarios in Chinese and English. According to these measurements, the model competes with the best open solutions and certain closed systems such as HeyGen, Kling Avatar 2.0, or OmniHuman-1.5.

The code, weights, and a technical report are available on GitHub and Hugging Face. The license is MIT, but its use is currently research-oriented, and generation requires a graphics card with substantial memory.