Chinese tech giant Meituan has open-sourced LongCat-Video, a unified framework for generating high-quality video from various inputs. According to the project's GitHub repository, the latest update on May 21, 2026, introduced LongCat-Video-Avatar-1.5, which significantly improves lip-sync accuracy for digital humans. The foundational model has 13.6B parameters.
LongCat-Video is designed as a single, powerful tool for developers. It handles text-to-video, image-to-video, and video continuation tasks within one framework. This release positions Meituan as a key player in the open-source AI video space, competing with models from both established tech companies and emerging startups.
The model is particularly noted for its ability to generate minutes-long videos without the common issues of color drift or quality degradation. This capability is crucial for applications beyond short social media clips, such as creating content for storytelling or virtual assistants.
What Can LongCat-Video Do?
LongCat-Video integrates multiple video generation tasks into a single architecture. It natively supports creating videos from text prompts, still images, or by extending existing video clips. Its pretraining on video-continuation tasks allows it to produce long-form content with consistent quality, a significant challenge for many models.The framework employs a coarse-to-fine generation strategy to efficiently produce 720p videos at 30 frames per second. This process is further accelerated by Block Sparse Attention, which boosts performance, especially at higher resolutions. This efficiency makes high-quality video generation more accessible without requiring massive computational resources.
How Does Avatar-1.5 Improve Lip Sync?
The LongCat-Video-Avatar-1.5 update focuses on creating more realistic digital humans. It achieves this by replacing the older Wav2Vec2 audio encoder with Whisper-Large-v3. This change allows for far more accurate and natural lip synchronization with audio inputs, a critical feature for creating believable virtual characters.
Beyond lip-sync, the v1.5 update introduces several key enhancements. Step distillation accelerates inference to just 8 steps, enabling faster video creation. Support for INT8 quantization also reduces VRAM usage, making the model runnable on less powerful hardware. The model now generalizes better to stylized domains like anime and animals.
Performance Against Competitors
LongCat-Video demonstrates performance comparable to leading proprietary and open-source models. The model was trained using a multi-reward Group Relative Policy Optimization (GRPO), a form of Reinforcement Learning from Human Feedback (RLHF). This advanced training helps it achieve high scores in visual quality and text alignment.Internal benchmarks show it holds its own against competitors. While some proprietary models may lead in certain metrics, LongCat-Video's open-source nature makes it a compelling alternative for developers and researchers. Meituan is also behind other major AI efforts, like the 1.6 trillion parameter LongCat-2.0 model.
Metric | LongCat-Video | PixVerse-V5 (Proprietary) | Veo3 (Proprietary) | Wan 2.2-T2V (Open Source) |
|---|---|---|---|---|
Overall Quality | 3.38 | 3.36 | 3.48 | 3.35 |
Text-Alignment | 3.76 | 3.81 | 3.99 | 3.70 |
Visual Quality | 3.25 | 3.13 | 3.23 | 3.26 |
Motion Quality | 3.74 | 3.81 | 3.86 | 3.78 |
# Activated Params | 13.6B | - | - | 14B |
