Meituan’s LongCat team released version 1.6 of its audio-driven avatar video model on November 15. The update targets a problem that becomes more obvious as a video runs longer: keeping a generated speaker’s face, movements and timing consistent from one segment to the next.
LongCat-Video-Avatar takes an audio track and uses it to animate an on-screen person. The project documentation describes single-speaker generation from audio and text or audio and a still image, as well as workflows for multiple audio inputs. This is distinct from a text-to-video model that must create both the scene and the spoken performance from a prompt.
Extending a video without compounding errors
The LongCat team says its update makes mouth movements less exaggerated and improves how the face, head and body move together. Its Hugging Face model card also says the changes reduce the gradual buildup of visual errors as the system uses earlier generated frames to continue a video. Meituan says the approach can produce videos longer than 10 minutes while keeping the speaker’s appearance and motion more consistent.
That duration is a claim from the model developer, not a guaranteed runtime or quality threshold for arbitrary recordings. A model that produces a convincing opening still has to preserve timing and identity late in a clip. For an instructional presenter, for example, changes in facial features or repeated gestures can distract from the content even when the words remain synchronized.
Downloadable weights, not a turnkey service
The underlying model files, known as weights, are downloadable from Hugging Face under the MIT License. Meituan hosts the inference code and examples on GitHub. The project documents installations with Python tools and GPU-related dependencies, and its version 1.6 examples include two-GPU configurations. The availability of model files does not mean a typical laptop can render long videos at practical speed.
On the model card, Meituan reports that version 1.6 placed second among six systems in a comparison based on 33,051 individual preference votes, behind OmniHuman-2.0 and ahead of version 1.5. That was the developer’s own evaluation. It does not tell a team what hardware it will need or whether the result will hold for a particular voice, visual style or recording condition.
Where the release changes the choice
For teams that can run a model themselves, downloadable weights give them a route to test avatar generation without designing around a particular hosted video service. They still need computing time, deployment work and permission to use the voices and visual material they provide. The model’s license covers the weights; it is not consent to use someone’s image or voice.
The practical question for a training or production team is whether the improvement extends to the full length of an ordinary project. A short approved sample may establish the framing and voice, while a test at the intended runtime can reveal late changes in appearance, lip timing or gesture repetition that a short sample would miss.



