Wan API: Open Video Generation by Alibaba

The Wan API gives developers Alibaba's open Wan video models on Atlas Cloud through one unified key. This page runs Wan 2.2, a foundational video model that pioneered a Mixture-of-Experts architecture for video diffusion, covering text to video, image to video, and video to video with cinematic motion control. It sits alongside 300+ other models on the same account, so you reach it with OpenAI-compatible endpoints and no separate setup.

Wan 2.2 ships in a few variants: the A14B models built for cinematic-quality text to video and image to video, and the lighter TI2V-5B model that combines both modes and runs at 720p on consumer GPUs. Atlas Cloud also offers turbo variants that trade a little fidelity for faster, lower-latency generation. Pick A14B for maximum quality, TI2V-5B for efficiency, and turbo for rapid iteration.

The Wan API covers text to video, image to video, and video to video, so you can start from a prompt, animate a still, or transform existing footage. It also supports first-to-last frame control, video extend to lengthen a clip, and upscaling to sharpen resolution. Each mode is a parameter or endpoint change, which keeps a mixed pipeline on one integration.

Mixture-of-Experts, or MoE, splits the denoising process across specialized expert models that each handle part of the work. Wan 2.2 was the first video diffusion model to use this design, which raises overall model capacity and detail without increasing inference cost per step. In practice that means richer scenes and more accurate motion at the same compute budget.

Wan 2.2 generates natively at 480p and 720p, with higher resolution available through upscaling to 1080p or 2K, and typical clip lengths in the short range of a few seconds per generation. The lighter TI2V-5B variant targets 720p at 24fps.

Yes. Wan 2.2 is released by Alibaba under the Apache 2.0 license, with model weights and inference code published on HuggingFace and GitHub, so you can download, modify, and self-host it. Running it locally needs a GPU with at least 16GB of VRAM, ideally more for the larger A14B models. Atlas Cloud hosts the Wan API so you can skip the hardware and scaling work entirely.

Yes. The Apache 2.0 license allows commercial use, modification, and redistribution, so video generated with Wan 2.2 can go into commercial projects. Review Atlas Cloud's terms of service for the specifics of your plan, and note the usual restrictions around generating content that depicts real, identifiable people without their consent.

The turbo variants apply an rCM sampler and 4-step distillation, which compress the denoising process into far fewer steps. This lowers latency and cost per clip while keeping strong visual fidelity, which makes turbo a good fit for iteration, batch runs, and prompt testing. Move to the standard A14B models when a final render needs maximum detail.

Generation is asynchronous: each request returns a prediction ID that you poll until the clip is ready, which fits queues and high-volume runs. Add exponential backoff and a retry on a 429 response, and use the turbo variants for drafts to keep throughput high. Contact support to raise concurrency limits as your workload grows.

Create an account on Atlas Cloud, generate an API key, and send a request to the Wan model with your prompt or input image through the OpenAI-compatible endpoint. Poll the prediction endpoint for the finished clip, then scale up as needed. Because the same key reaches 300+ models, you can test other video and image models without any extra setup.