A Chinese food delivery giant just released an open-source AI model that generates talking and singing avatars from photos and audio.
And apparently, delivering food wasn't enough.
Meet LongCat-Video-Avatar 1.6 from Meituan, one of China's largest platforms for food delivery, shopping, travel, and local services.
Yes, the company delivering your lunch is now building advanced generative video models.
And the results are surprisingly competitive.
WHAT'S NEW IN LONGCAT-VIDEO-AVATAR 1.6?
• 10+ minute videos: generate extended talking-avatar sequences while maintaining more consistent faces, clothing, and character appearance.
• Talking and singing avatars: animate characters using audio, including speech, music, and singing performances.
• Better lip synchronization: reduces exaggerated mouth movements, unnatural lip colors, and visual instability.
• Natural expressions: improved coordination between facial animation, head movement, body gestures, and audio.
• Multiple characters: supports single-person and multi-person audio-driven video generation.
• Eight-step inference: an optimized generation pipeline using DMD and OPSD LoRA adapters.
• Adjustable LoRA mixing: control how the two adapters influence different stages of the denoising process.
• INT8 quantization: optional lower-precision weights to reduce memory requirements.
• FlashAttention acceleration: FlashAttention-2 enabled by default, with optional FlashAttention-3 and xformers support.
• Open-source release: downloadable model weights and inference code under the MIT license.
HOW DOES IT COMPARE?
Meituan conducted a human-preference study comparing six avatar generation systems.
The evaluation included 33,051 valid votes across 15 model pairings.
Reported ranking:
1. OmniHuman-2.0
2. LongCat-Video-Avatar 1.6
3. LongCat-Video-Avatar 1.5
4. HeyGen
5. OmniHuman-1.5
6. Kling Avatar 2.0
LongCat 1.6 achieved a Bradley-Terry Elo rating of 1068.1.
That puts it ahead of several commercial avatar platforms in Meituan's evaluation.
But this is an internal benchmark organized by the model's developer, not an independently verified universal ranking.
THE TECHNICAL PART
LongCat-Video-Avatar 1.6 builds on Meituan's LongCat-Video foundation model.
The new release focuses on reducing the accumulation of errors during long video generation.
This matters because autoregressive video models can gradually distort faces, change clothing details, exaggerate expressions, or lose visual consistency as a sequence gets longer.
Version 1.6 attempts to reduce that drift while improving animation quality and inference speed.
The optimized pipeline uses eight denoising steps.
Two specialized LoRA adapters, DMD and OPSD, can be mixed during the initial steps before OPSD completes the remaining generation.
Developers can adjust the balance using mix_steps and mix_alpha.
CAN YOU RUN IT LOCALLY?
Yes, but don't expect a one-click consumer application.
The official examples use two CUDA GPUs, require downloading the foundational LongCat-Video model and additional avatar checkpoints, and involve a substantial Python inference environment.
INT8 quantization helps reduce memory consumption, but this remains a demanding video generation model.
Also, the demonstrated 10-minute consistency isn't a guarantee that every generated avatar will remain perfect throughout a long video.
THE BIGGER QUESTION
We're watching an increasingly strange AI market.
Food delivery platforms build foundation models.
Search companies build coding agents.
Social networks build video generators.
And everyone seems determined to become an AI infrastructure company.
But for developers, there's a genuinely interesting opportunity here.
Open avatar models could eventually power virtual presenters, educational content, digital characters, interactive entertainment, and automated video production.
Now I'm curious.
Is anyone actually implementing these open-source avatar models in real projects?
Or are we all just enthusiastically reading release notes and collecting Hugging Face links?
PROJECT & SOURCES
Model weights:
https://huggingface.co/meituan-longcat/LongCat-Video-Avatar-1.6
GitHub:
https://github.com/meituan-longcat/LongCat-Video
Official demos:
https://meigen-ai.github.io/LongCat-Video-Avatar-1.6-Page/
Technical report:
https://arxiv.org/abs/2611.18822
#LongCatVideoAvatar #AIVideo #OpenSourceAI
YouTube / ABVX
