Skip to content
AI.info

The Pulse

Meituan says open LongCat avatar model makes videos over 10 minutes

Meituan has released LongCat-Video-Avatar 1.6, an open model for audio-driven human video generation. The team says it can produce videos exceeding 10 minutes and ranked second among six systems in its human preference study.

Meituan says open LongCat avatar model makes videos over 10 minutes

Meituan has released an open model for generating audio-driven human videos that it says can run for more than 10 minutes. The LongCat team's model card says LongCat-Video-Avatar 1.6 reduces the accumulation of errors as a video is continued, while preserving visual quality, identity consistency and coherence over time.

The release was dated Nov. 15, 2026, in the team's GitHub repository. The model supports audio-driven video generation with one or multiple characters, and its weights are available on Hugging Face.

Longer avatar videos

The new version builds on LongCat-Video-Avatar 1.5. Its model card says the update improves synchronization between lips and audio, coordination of facial and body motion, and visual consistency during long-form generation.

Meituan says the update also reduces exaggerated or strained mouth movements, overly red lip coloration and instability around the mouth across frames. It describes changes to the generation pipeline and model execution as making inference faster and deployment more efficient.

The release repository describes a post-training pipeline that combines reinforcement learning, improved classifier-free guidance for DMD-based few-step distillation, long-video rollouts and on-policy distillation. The team credits that pipeline with enabling longer videos with improved visual quality and temporal consistency.

The model card provides examples for generating a single-person video from speech and text or from speech and an image. It also includes examples for continuing those videos and for generating and continuing multi-person animation from speech and an image.

Preference study

In a human preference study, the team compared the new model with its predecessor and four commercial systems: OmniHuman-2.0, HeyGen, OmniHuman-1.5 and Kling Avatar 2.0. The all-pairs study collected 33,051 valid individual votes across 15 model pairings.

The team reports that version 1.6 ranked second among the six systems, with a Bradley–Terry Elo score of 1068.1. It placed behind OmniHuman-2.0 and ahead of the other evaluated systems, including LongCat-Video-Avatar 1.5.

Weights and setup

The model weights are released under the MIT License. The license does not grant rights to use Meituan trademarks or patents.

The package uses the same base architecture as version 1.5 and includes two task-specific LoRA adapters, one for DMD and one for OPSD. An optional weight-only INT8 model is also included and can be enabled during inference.

The model card includes instructions for cloning the repository, installing dependencies and downloading both the foundational LongCat-Video model and the avatar weights. Its inference examples use separate scripts for single-person and multi-person animation.

The card cautions that the model has not been specifically designed or comprehensively evaluated for every downstream application. It asks developers to consider performance variations across languages and assess accuracy, safety and fairness before deployment in sensitive or high-risk settings.

Sources

  1. Meituan LongCat on Hugging Face
  2. Meituan LongCat on GitHub

Related stories

Explore

More articles