LongCat-Video-Avatar is an audio-driven avatar model built for long videos which was released end of December 2025.
It’s said to improve a lot over InfiniteTalk. Long sequences look more stable. Faces and motion hold together better too.
Built on the LongCat-Video base. It supports Audio-Text-to-Video, Audio-Text-Image-to-Video, and video continuation, all in one setup.
It’s open source under MIT license and claims top realism scores in EvalTalker tests. The tests used hundreds of people and multiple reviewers per video. But it appears you'd need around 32GB to run BF6.
Long videos stay clean. Cross-Chunk Latent Stitching helps stop blur and small errors from stacking up over time. Identity stays steady. Reference Skip Attention keeps characters consistent without stiff copy-like motion. It also handles more than one person and doesn’t lock you into short clips.
LongCat-Video-Avatar comes from the Meituan LongCat team. It builds on the LongCat-Video base model and focuses on avatar videos where audio and text drive motion, facial movement, and lip sync. You get people that move and speak in a way that feels more natural.
The model works across a few tasks. Audio-Text-to-Video. Audio-Text-Image-to-Video. Audio-based video continuation.
It’s meant for longer videos where faces don’t drift and motion doesn’t break. You’ll see smoother flow and steadier visuals as time goes on.
If you'd like to access this model, you can explore the following possibilities:
Weights