Wan 2.7 API

Wan 2.7 is Alibaba's new video generation suite, and its theme is control. Most video models still work like a slot machine: write a prompt, pull the lever, hope the shot lands. Wan 2.7 lets you pin down more of the result before you press go. You can fix the first and the last frame and let the model fill in the motion between them, hand it an existing clip and have it carry on, animate a still to your own soundtrack, or cast characters from photos, short videos and a voice sample.

Alibaba's own pitch is that the release moves creators "from executors to directors". The suite has four models: text-to-video, image-to-video, reference-to-video and an instruction-based video editor. All of them produce clips of 2 to 15 seconds (10 for references and edits) at 720p or 1080p, with sound generated alongside the picture. Unicorn API carries the three generation models: Wan 2.7 Text to Video, Wan 2.7 Image to Video and Wan 2.7 Reference to Video.

The launch post went up on 7 April 2026, a few days after its sibling, the Wan 2.7 Image model. What makes Wan 2.7 Video worth a look is less a leap in raw picture quality than the fact that keyframes, continuation, audio-driven animation and multi-reference casting now sit in one consistent family with one set of rules.

Who makes Wan 2.7, and is it open?

Wan is the video line of Tongyi Lab, the Alibaba group that also builds the Qwen language models. The family earned its reputation in the open-source world: Wan 2.1 and Wan 2.2 shipped their weights under the Apache 2.0 licence and became fixtures of local video pipelines. Wan 2.2 is still the version people download and run on their own GPUs, and Unicorn serves a quick variant of it as Wan 2.2 Fast.

The flagship line has been closed since Wan 2.5, which added native audio in September 2025, and Wan 2.7 continues that approach. There is no Wan 2.7 repository on Alibaba's Hugging Face organisation, no checkpoint on GitHub and no open licence. Some websites describe Wan 2.7 as an Apache 2.0 release, but no Alibaba channel says so, and the weights simply are not published. Access runs through Alibaba Cloud Model Studio, the wan.video site, the Qwen app and API providers such as Unicorn.

So if you need a Wan model you can fine-tune or run privately, Wan 2.2 remains the answer. If you want the newest Wan behaviour through an API, Wan 2.7 is it.

What changed from Wan 2.6

Wan 2.6, released in December 2025, already offered clips up to 15 seconds, native audio and multi-shot storytelling. Wan 2.7 keeps that envelope and builds out the controls around it.

First and last frame in one model. Wan 2.6's image-to-video model worked from a first frame only; to pin an ending you had to switch to a separate keyframe model. The Wan 2.7 image-to-video model accepts a start frame and an optional end frame, and generates the motion that connects them. That is what you need for product reveals, transformations, before-and-after shots and any edit where the next cut has to start on a known image.

Clip continuation. The same model can take an existing video of 2 to 10 seconds as its starting point and continue it, optionally steering towards a final frame. That makes it possible to grow a scene beyond one generation or to extend footage you already like instead of regenerating it.

Audio-driven animation. Image-to-video can also take a driving audio track (WAV or MP3, 2 to 30 seconds, up to 15 MB) alongside the first frame, or alongside first and last frames, and animate to it. Without an audio input, the model generates matching music or sound effects itself.

A broader reference model. Reference-to-video takes up to five reference images and videos in total, an optional first frame, and a short voice sample that sets the timbre of a character. Prompts now point at assets by type and number ("Image 1", "Video 2"), which reads more naturally than Wan 2.6's "character1" convention. Alibaba says the suite keeps up to five distinct characters consistent across videos and offers "over 50 refined emotional expressions".

Editing as a fourth model. Wan 2.7 adds an instruction-based editor that changes a clip's actions, dialogue, appearance, scene, style or camera work from a plain-language request, and re-syncs lips when dialogue is rewritten while keeping the original voice. This model is not on Unicorn, but it signals where the family is heading.

Storytelling in the prompt. Wan 2.6 had an explicit multi-shot switch. In Wan 2.7 text-to-video, you describe the shots and their timing in the prompt itself, and Alibaba says a narrative engine turns that into a paced storyboard with moves such as FPV drone passes and 360-degree orbits.

What did not change: the 15-second ceiling, the 720p and 1080p tiers, and the 30 fps output.

How Wan 2.7 works, as far as Alibaba says

Alibaba has published no technical report, parameter count, architecture details or training data for Wan 2.7, and no benchmark scores of its own. Figures circulating online for its size do not come from Alibaba.

A note on "thinking mode", which features heavily in launch coverage: the documented thinking switch belongs to Wan 2.7 Image, where it plans composition before rendering. The video APIs expose no such option. What they do have is prompt rewriting (prompt_extend), on by default, which expands short prompts into fuller scene descriptions before generation. It helps with terse prompts and can be switched off when you have written exactly what you want.

Wan 2.7 specs at a glance

Where Wan 2.7 shines

Shots with a fixed beginning and end. First-and-last-frame generation is the feature testers single out, with one review calling it the fix for a "persistent frustration" with AI video. Give it a closed jar and an open one, a bare room and a furnished one, a sketch and a finished render, and it produces the transition between them.

Growing a scene. Continuation lets you treat a good clip as a foundation rather than a one-off. Combined with a target last frame, you can steer a continuation into the next planned shot.

Characters from references. Reference-to-video takes photos and short clips of a person, pet or product and puts them into a new scene, with a voice sample when the character speaks. For series work, mascots and product films, that beats rewriting a description and hoping the face matches.

Visual polish at 1080p. Reviewers report sharper detail, more convincing skin and fabric, and steadier motion than Wan 2.6, along with better character consistency across shots.

Flexible lengths. Text-to-video and image-to-video accept every whole number of seconds from 2 to 15. You can make a 4-second loop or a 13-second spot instead of rounding to a fixed 5, 10 or 15-second block.

Weaknesses and limits

Speed. One platform that tested Wan 2.7 at launch measured about four minutes for a 5-second clip and called it significantly slower than competing models, without better output to show for the wait. Plan batch jobs accordingly.

Lip sync and dialogue. Lip sync is described as "functional but not frame-perfect". Short lines at a natural pace hold up, but fast delivery drifts, and when several characters talk at once, one tester found the voices "collapse into one dominant voice".

Physics and fast action. Reviewers find fast action and rapid physical dynamics weaker than in the best rival models, so energetic scenes need more re-rolls.

Arena standing. In Artificial Analysis's blind-vote text-to-video arena with audio, Wan 2.7 sits around tenth place, well behind ByteDance's Seedance 2.0. It is a capable generator, not the quality leader.

Prompting takes practice. Reviewers note a steeper learning curve than simpler tools: the model rewards structured prompts that spell out subject, action, camera and sound. Automatic prompt rewriting can also nudge results away from what you wrote.

Closed, with no draft tier. There are no weights, no fine-tuning and no 480p option for quick drafts (Wan 2.6 offered one upstream). Reference-to-video tops out at 10 seconds on Unicorn.

Prompting Wan 2.7: tips and example prompts

Write like a shot list: subject, action in order, camera, light, then sound. Keep spoken lines short and in quotation marks. Use the negative prompt for recurring artefacts (extra fingers, text overlays, flicker) rather than cramming exclusions into the main prompt.

For text to video, a close shot with clear action and sound:

Close-up at a night market: a street cook tosses noodles in a flaming wok, sparks and steam rising under warm neon light. She glances at the camera and grins. Sizzling oil, clanging metal, crowd chatter.

For multiple shots in one clip, give each shot its timing:

Shot 1 (0-4 s): wide shot of a lighthouse on a cliff at dusk, waves breaking below. Shot 2 (4-8 s): inside, an old keeper climbs the spiral stairs with a lantern. Shot 3 (8-12 s): the beam sweeps across the dark sea. Wind, surf, footsteps on iron steps.

For first and last frame, describe only the change between the two images:

The flower bud slowly opens into full bloom as morning light moves across the petals. Locked-off macro shot, no camera movement.

For reference to video, name each asset by type and number:

The woman from Image 1 walks into the bakery from Image 2, smiles at the counter and says, "Two croissants, please." Medium shot, soft window light.

Two habits pay off. Draft at 720p and short durations, then re-render the winner at 1080p with the same seed. And if a short prompt drifts, write a fuller one and turn prompt_extend off so the model follows your wording.

How Wan 2.7 compares with other video models on Unicorn

Against Wan 2.6. Wan 2.6 Text To Video works only in fixed 5, 10 and 15-second blocks, and its image-to-video model has no last frame, continuation or driving audio. For anything that needs control, Wan 2.7 is the upgrade.

Against Seedance 2.0. Seedance 2.0 scores clearly higher with human voters and is the pick when raw picture quality matters most. It tops out at 720p rather than 1080p, but offers a 480p tier for quick drafts.

Against Kling 3.0. Kling 3.0 offers standard, pro and 4K quality tiers with optional sound, so it suits jobs that need a resolution above 1080p. Wan 2.7 is the better fit for keyframe control and reference casting.

Against Wan 2.2 Fast. The open-weights Wan 2.2, served here as a fast image-to-video model, is fine for quick silent animation at up to 720p. It has no audio, no references, no continuation and no 1080p.

Choosing between the three Wan 2.7 models

The three models share a rulebook, so the choice comes down to what you already have. Start with text to video when the idea exists only in words; it is also where the shot-by-shot timing in the prompt does the most work. Use image to video when a frame is already approved: give it a first frame for a straightforward animation, add a last frame when the shot must land on a known image, pass a clip to continue a scene, or add a driving audio track when the motion should follow music or a voice. Use reference to video when a particular person, pet or product has to appear, especially with a voice sample for a speaking part.

Watch the defaults when sending requests. Text-to-video and reference-to-video default to 5 seconds at 1080p, while image-to-video defaults to 2 seconds at 720p, so set duration and resolution explicitly to get the clip you have in mind.

Getting started

Open the Wan 2.7 text-to-video playground, or the image-to-video playground to upload a first frame, choose duration and resolution, and generate. Reference-to-video lives in the same studio under wan-2-7-r2v.

For code, create a key on the API keys page and submit a job to /v1/jobs with the model id and a prompt, plus duration, resolution and any image, audio or reference inputs. The job returns a video URL when it finishes; the API docs cover authentication, polling and webhooks.

Verdict

Wan 2.7 does not top the quality charts, and it is slow and closed. But it turns Alibaba's video line into a real toolkit: start and end frames, continuation, audio-driven animation and multi-reference casting with voices, all under one set of rules. If you know the shot you want and need the model to hit it, Wan 2.7 gives you more ways to get there than any Wan before it.