Back Home

GitHub Repo

Meituan Releases LongCat-Video-Avatar 1.6, Opens Weights and Improves Long-Video Generation Stability

Meituan announced LongCat-Video-Avatar 1.6 on November 15, an audio-driven human video model with MIT-licensed weights and inference code. The new version combines post-training with a dual-LoRA distillation setup. The team says it improves consistency in videos longer than ten minutes, though actual quality and deployment costs still need independent validation.

Capwiuejooh · CC BY-SA 4.0 · Image source
zh-Hant

The Meituan LongCat team announced LongCat-Video-Avatar 1.6 on November 15, releasing weights and inference code for audio-driven human video generation. The model supports single-person and multi-person scenes, can generate video from speech, text, or reference images, and can continue existing clips. This update focuses on more natural lip motion and reducing errors that accumulate as videos are extended over time. Official release notes

The main technical changes are in the post-training pipeline. The team combines reinforcement learning, improved classifier-free guidance for few-step DMD distillation, long-video rollouts, and on-policy distillation. It says this helps maintain image quality, character identity, and temporal consistency in generations longer than ten minutes. These improvements matter directly for long-form virtual character content: accurate lip motion in a short clip does not establish that the same face and coherent movement will persist through repeated continuations. The ten-minute capability should currently be treated as a result reported by the team. Technical change notes

For deployment, version 1.6 retains the 1.5 architecture and uses two LoRAs, DMD and OPSD. The official example uses eight denoising steps: DMD for the first four and OPSD for the last four. Developers can adjust the number of steps allocated to each and their weights to test structural stability. The weight package also provides an INT8 DiT with weights only, and notes that INT8 and FP8 modes cannot be enabled at the same time. Model card and inference settings

For evaluation, the team conducted pairwise comparisons across six systems and collected 33,051 valid human preference votes. LongCat-Video-Avatar 1.6 scored a Bradley–Terry Elo of 1068.1, ranking second in that test behind OmniHuman-2.0. This was a preference study conducted by the publisher, and does not establish a quality ranking across all languages, people, or video lengths. Engineering teams should next test lip synchronization with target audio, drift across continuations, GPU memory usage, and generation time per minute. The MIT-licensed weights make it easier to run these validations locally. Evaluation and licensing details

Sources

  1. LongCat-Video 官方儲存庫與 2026 年 11 月 15 日發布紀錄
  2. LongCat-Video-Avatar-1.6 官方模型卡