What is LongCat-Video?
LongCat-Video was released on 25 October 2025 by Meituan's LongCat team. It is a 13.6-billion-parameter diffusion transformer that handles text-to-video, image-to-video and video continuation in one model, and because it was pretrained on continuation it can extend a clip repeatedly into a video minutes long. It generates 720p at 30 fps by refining coarse output in time and resolution, with block sparse attention to cut the cost at high resolution.
The weights are MIT-licensed and there is no audio. The code also includes an interactive generation mode and a Streamlit demo. The GitHub repository has over 9,000 stars and was last updated in May 2026.
Is LongCat-Video worth it?
LongCat-Video's one distinctive skill is length. Because it was trained to continue existing video, it can extend a clip again and again into minutes of footage while keeping the scene coherent, which models built only for single short clips cannot do.
On quality, Meituan's own tests place it in the 2025 field. In its internal text-to-video ratings it scores 3.38 overall against 3.48 for Veo 3 and 3.35 for Wan 2.2, and in image-to-video it trails Wan 2.2 at 3.17 against 3.26. It does not appear on Artificial Analysis's current text-to-video board, where every ranked model generates sound.
Its licence and price are the attractions. MIT terms carry no revenue cap or regional limits, and fal charges $0.04 per generated second at 720p, so a minute of video costs $2.40.
Long videos cost time, though. Each extension is another generation run, the model card gives no figures for how quality holds over several minutes, and it publishes no memory requirement. Development has slowed: the weights date from October 2025 and the repository's last push was in May 2026.
Where it earns its place
A property-listing site adding 2,000 homes a month
Make one-minute silent walk-through style videos from listing photos
Start from a photo with image-to-video and extend with continuation on fal at $0.04 a second: $2.40 per one-minute video, about $4,800 a month, with music added afterwards.
A digital-signage company serving shops in several countries
Produce long ambient background loops for in-store screens
Self-host the MIT weights and chain continuation runs into multi-minute 720p, 30 fps sequences, with no licence fee and no limits on where the screens are.
Before you commit
- Setup
- Create a conda environment, install PyTorch, FlashAttention 2 and the requirements, download the weights with huggingface-cli and run the single- or multi-GPU script for each task.
- Integrations
- Native PyTorch scripts and a Streamlit demo, CacheDiT acceleration of about 1.7x, and fal's hosted text-to-video and image-to-video endpoints.
- Data and privacy
- MIT weights run offline, so footage stays on your hardware; training data is not detailed on the model card.
- Cost at scale
- fal bills generated seconds at 30 fps, so 100 one-minute videos cost $240; self-hosting GPU requirements are not published.
- Limits
- No audio; 720p ceiling; quality evidence is Meituan's internal ratings; last code update May 2026.
What’s good
- MIT licence with no revenue or regional limits
- Continues clips into minutes-long video
- 720p at 30 fps
- fal API at $0.04 per second
What’s not
- No audio
- Internal scores trail Veo 3 and, for image-to-video, Wan 2.2
- Not on Artificial Analysis's current video board
- No published memory requirements
Best for
Developers who need long, silent 720p sequences such as walk-throughs, ambient loops or B-roll under an unrestricted licence.
LongCat-Video alternatives
LongCat-Video: common questions
- Is LongCat-Video worth it?
- LongCat-Video is the open model for footage that has to keep going, with an MIT licence and a cheap API, but it is silent and its quality is a year old. We rate it 3.5 out of 5.
- How much does LongCat-Video cost?
- Free weights under MIT. fal charges $0.04 per generated second at 720p and 30 fps.
- What are the downsides of LongCat-Video?
- No audio; internal scores trail Veo 3 and, for image-to-video, Wan 2.2; not on Artificial Analysis's current video board; no published memory requirements.
- Who is LongCat-Video best for?
- Developers who need long, silent 720p sequences such as walk-throughs, ambient loops or B-roll under an unrestricted licence.
- How much does LongCat-Video cost for a team or in production?
- fal bills generated seconds at 30 fps, so 100 one-minute videos cost $240; self-hosting GPU requirements are not published.
- Is LongCat-Video safe for company data?
- MIT weights run offline, so footage stays on your hardware; training data is not detailed on the model card.
- What are the best LongCat-Video alternatives?
- The closest alternatives we have reviewed are HunyuanVideo 1.5, Kandinsky 6.0 Video, LTX-2.5 and Wan 3.0.
Answers checked Nov 7, 2026.