Alibaba: Wan 3.0 Prime

Wan 3.0 Prime is a video generation model from Alibaba, listed on api14 under the model id alibaba/wan-3.0-prime. It is a fast-mode variant of Alibaba's Wan 3.0, and its stated purpose is text-to-video and first-frame image-to-video generation. If you are already familiar with the Wan line, treat this as the quicker sibling rather than a different family of model.

The model takes text and image inputs and produces video as its only output. Because it accepts a first frame image, you can drive a clip from a still you supply, or generate entirely from a written prompt. Output can be requested at 480p, 720p or 1080p, in aspect ratios of 16:9, 4:3, 1:1, 3:4 and 9:16, which covers landscape, square and vertical formats. Clip durations run from 2 to 30 seconds in one second steps, so you can ask for short loops or longer sequences without stitching. Audio generation is available as well, which means a single request can return video with sound rather than a silent track. No context length is published for this model, so plan around the video parameters rather than a token budget.

Those capabilities point to a fairly specific set of jobs. Vertical 9:16 output at short durations suits social and mobile placements, while 16:9 at higher resolutions fits web, presentation and broadcast style content. First-frame image-to-video is useful when you have approved artwork, a product shot or a storyboard panel and want motion derived from it rather than a fresh interpretation. The 2 to 30 second range works for teasers, animated stills, background plates and loopable clips. Audio generation reduces the need for a separate sound pass on simple pieces. Square and 4:3 ratios help when the same asset has to run across more than one channel.

You call it through api14's OpenAI-compatible API using the model id alibaba/wan-3.0-prime, so existing OpenAI style client code needs little more than a changed model string. The request parameters it supports are max_tokens, presence_penalty, response_format, seed, temperature and top_p. Seed is the one to reach for when you need reproducible output across runs, and temperature and top_p give you control over variation between generations. Note that max_tokens is supported as a parameter even though the model's output is video, and the published maxOutput value is 0, so do not rely on it to cap clip length. Duration, resolution and aspect ratio are set through the video specific fields rather than the token parameters.

There is no published context length, and the model is video only, so it is not a fit for text chat, classification, summarisation or any task that expects a text or structured response. If you need those, a text model is the better choice. Within video work, the main constraint to plan around is duration: 30 seconds is the ceiling, so longer pieces need to be generated in segments and joined afterwards.