Meituan releases LongCat Video Avatar 1.6 for videos over 10 minutes

Meituan’s LongCat team announced LongCat-Video-Avatar 1.6 on Nov. 15. The model generates video of people speaking in time with recorded audio. The developers say the new version can make videos longer than 10 minutes with less drift in the character’s face and mouth. The project merged the code for 1.6 into its main branch on Nov. 15. Its weights and model card had been uploaded to Hugging Face on Nov. 13.

Teams making extended explainers or dialogue scenes can download the new weights and evaluate them in their own production environment. The model works from audio with a reference image or a text description, and it can continue an existing clip. The previous version could already handle more than one speaker. Version 1.6 focuses on appearance across longer sequences and on how the mouth and face move during speech.

Keeping one scene consistent across clips

LongCat assembles longer videos from shorter clips. It carries some of the previous clip’s final frames forward, then appends the new frames it generates. A small change around the mouth or in the texture of a face can become more visible each time the model uses its earlier output as a reference.

The developers say they refined version 1.6 by training it in situations where it had to build on clips it had generated itself. They published side-by-side examples using the same input for both releases. Ten minutes refers to the duration of the resulting video. The model card’s execution examples use two GPUs. The materials give no minimum GPU memory figure or wall-clock time for rendering a video of that length.

What viewers preferred and what reviewers caught

Meituan’s comparison used 508 image-and-audio inputs. Reviewers saw two videos made from the same input and indicated which they preferred. The study compared six models in all 15 pairings and recorded 33,051 valid votes. Some systems failed to produce a usable clip on particular inputs, so the number of comparisons varied by pairing. Version 1.6 placed second, behind OmniHuman 2.0.

Reviewers could also mark both results as good or both as bad. The researchers counted those judgments as ties, crediting each contender with half a win. Under that method, version 1.6 had a 55.6% win rate over version 1.5 in 2,535 valid votes between the two.

Meituan ran a separate assessment of how human the output appeared. Results improved for realistic reference images and declined for animated ones. Experts also watched clips at half speed to spot lip-sync mistakes. They marked 82.5% of the version 1.6 clips as having a mismatch under that check, compared with 83.7% for version 1.5. That half-speed check was run on the assessment material, and the team did not report a separate rate for videos over 10 minutes.

The weights carry an MIT license. The listed files for the main model’s weights add up to about 31.7 GB in the standard set and about 15.9 GB in the optional 8-bit set. Those are file sizes. The amount of GPU memory needed to generate a clip also depends on the material being processed.

Hugging Face’s model page showed no hosted inference provider at the time of checking. The entire repository is listed at 77.6 GB and includes both sets of weights, supporting files and example videos.

The developers say they have not evaluated the model for every possible use. The model card identifies accuracy, safety and fairness as areas to assess before using it in sensitive settings. Downloadable weights give teams control over where the software runs. Computing costs and rights to the images and recordings used to make a video still depend on the production.