Meituan LongCat open-sources LongCat-Video-Avatar-1.6 for audio-driven human video generation
Meituan LongCat released LongCat-Video-Avatar-1.6, an open-source audio-driven human video generation model, on Hugging Face under an MIT license. It improves lip sync, supports videos over 10 minutes, and ranks second in a 33,051-vote human preference arena with an Elo of 1068.1.
China context
- Original name
- 美团龙猫
- Outside China
- Open weights · huggingface.co
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Builders can integrate the model via Hugging Face and use the provided LoRA mixing and INT8 options for deployment.
- For investors
- Meituan's open-source release competes with commercial avatar services, potentially pressuring pricing in the AI avatar market.
LongCat-Video-Avatar-1.6 is the latest upgrade to Meituan LongCat's open-source framework for audio-driven human video generation. It builds on version 1.5 with more natural lip animation, stable long-video generation exceeding 10 minutes, and faster end-to-end inference. The model is available on Hugging Face and ModelScope under an MIT license.
The model uses the same base architecture as 1.5 with two task-specific LoRA adapters (DMD and OPSD) and an optional INT8 DiT. Inference optimizations include FlashAttention-2 by default, with options for FlashAttention-3 or xformers, and a mixing strategy for the first denoising steps.
Developers outside China can now use an open-weights model that outperforms HeyGen and Kling Avatar 2.0 in human preference, reducing reliance on commercial avatar APIs.
The MIT license and open weights allow commercial use without licensing fees, potentially lowering costs for avatar video generation compared to proprietary services.
Watch for whether Meituan LongCat releases further versions or technical report details that confirm the claimed 10-minute stability and inference speedups in independent benchmarks.