Alibaba models and APIs

About Alibaba

Alibaba reaches this catalogue through two lines: the Wan video families and the image families published as Qwen Image and Z-Image. Together they account for more model families here than any other creator, and they cover every media task the platform exposes apart from chat and audio.

On video, Wan 3.0 and its Pro, Prime and Pro Prime tiers each publish text-to-video, image-to-video and reference-to-video side by side, returning 2 to 30 seconds with audio in the same file; the Pro tiers add resolutions up to 4K. Wan 2.6 and 2.5 return fixed clip lengths instead of a range, Wan 2.2 Animate transfers motion from a driving video onto a character image, and the LoRA variants let you attach your own weights. On images, Qwen Image 2.1 publishes text-to-image and reference editing as a matched pair, across fifteen aspect ratios and three resolution tiers, and Qwen Image 2.1 LoRA runs those same two endpoints with up to three of your own LoRA weights loaded per request, each passed as a URL with its own strength. Qwen Image 2512 LoRA does the same on the open Qwen-Image-2512 and its Edit-2511 editor, the base with the larger community LoRA library, and Z-Image Turbo LoRA puts your own weights on the 8-step Z-Image Turbo, with image-to-image and a strength control. Qwen Image 3.0 and 3.0 Pro handle text-to-image and multi-reference editing, Qwen Image 2 is the lighter editor, and Z-Image turns a tight brief into a photoreal frame in about a second.

Pick the Wan tier by resolution and clip length first, because that is what narrows the list fastest. Choose between the Qwen Image and Z-Image lines by whether you need legible type and multi-reference editing or raw speed, and reach for a LoRA family when the look you are after already lives in weights you hold rather than in a prompt. Pick the one whose base your weights were trained on: LoRAs for Qwen Image 2.1, Qwen-Image-2512 and Z-Image Turbo do not carry over from one to another. Uncensored variants are published as their own families rather than as a flag on the base model, so they are separate pages with separate schemas.