Meituan’s LongCat-Video Sets New Benchmark for Open-Source Text-to-Video

Meituan has officially open-sourced LongCat-Video, a powerful unified model that merges text-to-video, image-to-video, and video continuation into a single system — a major step forward in open AI video generation.

A New Era for Open-Source Video AI

LongCat-Video stands out for delivering minutes-long, high-quality videos without the typical color drift or quality degradation seen in other models. With 13.6 billion parameters, it represents one of the strongest open-source DiT-based (Diffusion Transformer) architectures in existence, optimized for both performance and accessibility.

This release gives developers and creators full access to production-level video generation — a feat previously dominated by closed systems such as OpenAI’s Sora or Google’s Veo.

Key Features and Technical Highlights

• Unified multitask architecture: Combines text-to-video, image-to-video, and video continuation under one robust base model.
• C2F pipeline + Block Sparse Attention: Enables fast 720p, 30fps rendering within minutes, maintaining visual consistency and smooth motion.
• Open weights & reproducibility: Designed for researchers to customize and extend; no proprietary APIs or cloud restrictions.
• Industry-standout quality: Capable of generating cinematic-level visuals with stable lighting, fluid motion, and accurate temporal alignment.

With its 13.6B parameter configuration and fine-tuned control layers, LongCat-Video pushes open research toward parity with leading commercial video models — and in some cases, surpasses them in flexibility and scaling potential.

Links & Resources

• GitHub: github.com/meituan-longcat/LongCat-Video
• Hugging Face: huggingface.co/meituan-longcat/LongCat-Video
• Project Page: meituan-longcat.github.io/LongCat-Video

Why It Matters

LongCat-Video’s open release lowers the barrier for independent creators, AI filmmakers, and research labs, enabling them to build, fine-tune, and deploy video models locally. The blend of speed, stability, and open access positions it as one of 2025’s most significant open-source AI video milestones.