Meituan Open-Sources LongCat-Video: A Single 13.6B Model for T2V, I2V, and Video Continuation
> Research subject: meituan-longcat/LongCat-Video
> Captured: 2026-10-03 | Stars that day: 43
> License: MIT | Language: Python
---
One Model, Three Tasks, One Architecture
A long-standing problem in video generation: Text-to-Video (T2V), Image-to-Video (I2V), and Video-Continuation have historically required three separate models, or one model with three sets of weights. Training silos, inference switching—huge engineering overhead.
LongCat-Video's first selling point: one model natively supports all three tasks, no task-specific fine-tuning, no weight switching. Give it text, it generates video; give it an image plus text, it generates video starting from that image; give it a video, it continues it.
That sounds like an engineering optimization, but it implies the model learned "spatiotemporal continuity of video" as a底层 (underlying) grammar during pretraining, rather than bolting it on with later task adaptation.
The Color-Drift Problem in Long Video Generation
Most video generation models share a flaw: the longer the video, the more colors drift. A frame that looks fine at second 5 may have a completely different tone by second 30, as if a filter were applied. The cause: no global color-consistency constraint when generating frame by frame.
LongCat-Video's solution is to include Video-Continuation as a pretraining task. Given the first N seconds of a video, the model must generate the next M seconds. This forces the model to learn "inherit visual features from existing frames" instead of generating each segment from scratch.
The result: the model can generate minute-long videos without color drift or quality degradation. It sounds like a small detail, but "can it generate 1+ minute videos" vs "10+ second videos" are entirely different application scenarios.
Block Sparse Attention: Solving the Compute Bottleneck of Long-Video Inference
The biggest obstacle to long video generation isn't model capability—it's compute. A 720p, 30fps, 1-minute video has 1,800 frames. With standard Dense Attention, every frame attends to all others—O(N²) complexity. 720p long videos simply can't run.
LongCat-Video uses Block Sparse Attention: split video frames into blocks, Dense Attention within blocks, Sparse Attention across blocks (attending only to key frames). Complexity drops from O(N²) to roughly O(N·log N).
Combined with a coarse-to-fine generation strategy (low-resolution draft first, then high-resolution refinement), LongCat-Video can produce 720p 30fps video in minutes of inference time. "Minutes" refers to inference time, not video length—wait a few minutes and get a usable long video.
Multi-Reward GRPO: Teaching the Model What Humans Find Good-Looking
The awkward situation for video generation models: objective metrics (PSNR, SSIM) look great, but humans say "something's off." Objective metrics measure pixel-level similarity, while humans judge video on multiple subjective dimensions—aesthetics, motion smoothness, semantic consistency, physical plausibility.
LongCat-Video aligns with human preference using multi-reward GRPO (Group Relative Policy Optimization):
1. Generate a group of candidate videos 2. Score them with multiple reward models (aesthetics, motion consistency, semantic matching, etc.) 3. Rank relatively within the group (relative policy), rather than using absolute scores 4. Update the policy with the ranking signal
This is more stable than traditional PPO because relative ranking eliminates reward-scale differences. Combining multiple reward models also avoids the trap of "optimizing one metric while degrading others."
13.6B Parameters: A Noteworthy Scale Choice
LongCat-Video has 13.6B parameters—a number worth pondering.
Current video generation models are polarized: small 1–3B lightweight T2V models, and 30B+ flagships. 13.6B is a middle ground—large enough to support a unified multi-task architecture and long-video generation, small enough to run inference on reasonable GPU setups.
Meituan's logic is likely: the commercial bottleneck for video generation isn't model capability, it's inference cost. A 30B model, even if better, is unusable for ordinary users if a single generation takes 30 minutes. 13.6B balances "good enough quality" with "affordable inference."
Why Is Meituan Building a Video Generation Model?
It looks like a leap, but Meituan's core businesses—local life services, dining, retail—need massive amounts of visual content: dish showcases, store environments, product demos, ad creatives. Demand is huge; traditional video production is expensive and slow.
A model that generates minute-long 720p video in minutes means: merchants can generate dish showcase videos from text descriptions, operations teams can batch-produce ad materials, and user reviews can upgrade from text to short video.
This differs from OpenAI's Sora—Sora is a general-purpose video generation platform, while LongCat-Video is closer to video generation infrastructure tailored to local-life scenarios. Meituan's Technical Report mentions "first step toward world models," hinting at grander plans ahead.
What the Open-Source Release Means
LongCat-Video's release under MIT license has several industry implications:
1. Lower engineering barriers. Techniques like Block Sparse Attention and multi-reward GRPO go from papers to runnable code. 2. Another option for open video generation models. With Sora closed and Stable Video Diffusion limited, LongCat-Video is a 13.6B-parameter, commercially usable, near-commercial-quality open option. 3. Validation that long video generation is feasible in open source. Most open models previously managed only 5–10 second clips; LongCat-Video proves minute-level long videos are achievable in the open-source community.
An Underestimated Detail: Video-Continuation as a Pretraining Task
Most models pretrain on T2V or I2V; LongCat-Video includes Video-Continuation in pretraining itself (not post-hoc fine-tuning). The significance:
- T2V teaches the model to "imagine visuals from text"
- I2V teaches it to "imagine motion from a static frame"
- Video-Continuation teaches it to "understand spatiotemporal regularities from existing video"
By placing Video-Continuation in pretraining, Meituan embedded "spatiotemporal continuity" into the model's underlying grammar. That's why LongCat-Video can generate long videos without color drift—it isn't "generating" long videos, it's "continuing" them.
---
Repository: https://github.com/meituan-longcat/LongCat-Video Technical Report: see the Technical Report link in the repo README Model weights: available on HuggingFace
---
> Daily GitHub Trending deep dive · Entry #83