Upload a photo and an audio file, and LongCat Avatar 1.5 turns them into a talking, singing, or animated video, no editing skills required.
Try LongCat Video Avatar For Free
LongCat Avatar 1.5 does more than move a mouth to match audio. It assembles a full talking, singing, or animated video from just a photo and a sound file.
One model handles 3 jobs at once. Audio Text to Video constructs a speaking character from just a script and a voice track. Audio Text Image to Video animates a reference photo you supply, and Video Continuation extends an existing clip to produce longer speaking shots.
LongCat Avatar 1.5 uses Whisper Large to read the audio and time each mouth movement to the actual speech. This keeps lip sync tight and natural, even across longer clips with changing tone and pace.
The tool can take 2 separate audio streams and sync each one to a different person in the frame. This lets both speakers talk, react, and take turns naturally in the same scene. It's a solid fit for interviews, duets, or any clip where 2 people share the frame.
It runs on just 8 generation steps with INT8 quantization, cutting down processing time without a real drop in quality. This means faster output, so you're not stuck waiting long for your video to render.
Running a model on someone else's server means accepting their limits, costs, and downtime. Here's why teams choose to self-host LongCat Avatar 1.5 instead.
LongCat is released under the MIT license, so teams can use, modify, and deploy it without paying license fees or asking for permission. This is different from tools that call themselves free but still charge per generation or lock features behind a paywall.
Start Creating Free
The tool keeps a character's face, features, and identity consistent even in long speaking or singing shots. This means no drifting or subtle changes as the clip runs longer, so the same person stays recognizable from start to finish.
Start Creating Free
This model does much more than a person sitting and talking at a desk. Anime characters, animals, multi-person scenes, and object-handling shots also come out steady.
Start Creating Free
Teams weighing avatar tools usually care about cost, control, and how the output actually looks. Here's how LongCat Avatar 1.5 compares against the hosted options on those points:
| Platform | License | Hardware Needed | Resolution | Users |
|---|---|---|---|---|
| LongCat Avatar 1.5 | MIT, free commercially | 40GB GPU minimum | 480P and 720P | ML engineers |
| Magiclight AI | Rights on paid plans | None, runs in a browser | 1080p, up to 50 minutes | Story creators |
| Synthesia | Minute-based subscription | None, runs in a browser | Studio quality | Enterprise teams |
| Mango AI | Subscription | None, browser or phone | Standard video | Marketers |
| HeyGen | Subscription | None, runs in a browser | Studio quality | Growth teams |
Setting up LongCat Avatar 1.5 takes just a few steps, whether you run it locally or through a hosted option.
Use a GPU with enough VRAM for the model and its audio encoder. The official setup supports INT8 mode to reduce memory use, with 40GB-class GPUs used for this configuration.
Download both the LongCat-Video base model and LongCat-Video-Avatar-1.5 weights from Hugging Face. The Avatar 1.5 setup uses the base model together with its own model weights.
Provide the audio and, for image-based generation, a reference image and text prompt. Longer, more detailed prompts are recommended because they can improve consistency and natural results.
For Avatar 1.5, use the 8-step distilled mode and choose 480P or 720P resolution. Audio CFG is recommended around 3–5 when using the standard sampling controls, while INT8 mode can reduce VRAM use.
ML Engineer
Our per-minute avatar bill was climbing faster than usage. Self-hosting LongCat moved that cost onto hardware we already owned and had sitting idle overnight.
Startup Technical Lead
The MIT license on LongCat is why we picked it over three better-documented options. Constructing a product on weights nobody can revoke was worth every hour of setup pain.
Ecommerce Video Creator
We generate product narrations in batches rather than one at a time. LongCat Video Avatar 1.5 runs them overnight, and nobody has to appear on camera.
Animation Studio Lead
Most avatar models fall apart the moment you feed them a stylized character. LongCat handles our anime designs without the face melting halfway through a shot.
An open-weight, audio-driven avatar video model from Meituan's LongCat team, released in May 2026. It is constructed on the 13.6 billion-parameter LongCat-Video foundation model and is officially written with hyphens as LongCat-Video-Avatar 1.5.
Yes, under an MIT license, which is about as permissive as open weights get. You can ship client work, produce a product on it, or run it at scale without paying a license fee or reporting usage to anyone.
A 40GB GPU is the realistic floor. One hands-on test ran it on an A800 with INT8 quantization and the eight-step distill. That still measured around 44 seconds of GPU compute per second of finished video.
Anime characters, animals, and multi-person scenes all work, including shots where somebody handles an object. Singing and acting hold up too, which is unusual for a model created around speech.
Raise the audio CFG value, which works best somewhere between 3 and 5. Write longer, more descriptive prompts too, because short ones give the model less to keep consistent across a clip.
Whenever nobody on the team runs GPUs for a living. Setup is genuinely hard, compute is slow, and a subscription tool beats it outright unless you already have the hardware and engineering time.
Clone the model and run one clip through it tonight. Bill yourself for the GPU time before anyone signs a subscription.