"LongCat AI" is one name for several different things, and mixing them up is the quickest way to download the wrong weights or follow the wrong tutorial. LongCat is an open-source AI generation framework developed by Meituan. Under that banner sit at least three generation models — a video model, an image model, and an audio-driven avatar model — plus a separately named language model that has nothing to do with video generation. This page maps the lineup so you can pick the right model before you spend an afternoon on the wrong one.
| Model | What it generates | What it is for |
|---|---|---|
| LongCat Video | Video, up to 961 frames | Text-to-video, image-to-video, and video continuation |
| LongCat Image | Still images | Image generation that runs on modest VRAM |
| LongCat Video Avatar 1.5 | Talking-head video | Audio-driven lip-synced avatars from a portrait + voice clip |
| LongCat-Flash | Text | A separate 560B-parameter language model for reasoning, not video |
The critical separation: LongCat-Flash is a language model, not a video model. If you want reasoning over text, that is the name to reach for. If you want video or images, you are dealing with the generation models above.
LongCat Video was released by Meituan in September 2025 under an MIT license. It is a 13.6 billion parameter Dense Transformer built around one idea: temporal coherence over long sequences, where most short-clip models start to drift:
The 13.6B figure and the frame ceiling come straight from the model's published description; the differentiator worth remembering is continuation, which is what makes minutes-long shots practical instead of a slideshow of glued clips.
LongCat Image is the still-image sibling. The relevant practical point is the opposite of LongCat Video's: where the video model is heavy and long-form, the image model is positioned to run on lower-VRAM hardware. It is not the same checkpoint as the video model — download the image weights for image work, and do not assume one download covers both.
LongCat Video Avatar 1.5 is the audio-driven path. Feed it a portrait and a short WAV, and it generates a lip-synced talking head. Its audio encoder and inference path differ from the base video model, so treat it as its own tool rather than a setting on LongCat Video. This is the model behind faceless presenters and recurring virtual hosts.
Open weights are not a commercial license. "MIT" and "open source" describe how you may use and modify the code and weights; they do not automatically grant rights to the portrait in your avatar, the music in your video, or a third-party checkpoint you layered on top. Review the license of each model you download and the rights attached to every input before you publish or sell the result.
Explore the lineup from the LongCat Video homepage.
LongCat is an open-source AI generation framework developed by Meituan. It includes a video model (LongCat Video), an image model (LongCat Image), and an audio-driven avatar model (LongCat Video Avatar 1.5).
No. LongCat-Flash is a separate 560-billion-parameter language model for text reasoning. It has nothing to do with LongCat Video or image generation, despite sharing the LongCat name.
LongCat Video is a 13.6 billion parameter Dense Transformer. It generates up to 961 frames at 480p or 720p and supports text-to-video, image-to-video, and video continuation.
LongCat Video generates motion — clips and continued sequences. LongCat Image generates still images and is positioned to run on lower-VRAM hardware. They are separate checkpoints; one download does not cover both.
Not by itself. Open weights describe how you may use the model, not the rights to your inputs or to third-party checkpoints. Review each model license and the rights to every portrait, audio, and dependency before commercial use.
Video, image, avatar, or reasoning — pick the correct checkpoint and skip the wrong-tutorial rabbit hole.