Wan 2.7: The Complete Guide (2026)

Wan 2.7 is Alibaba Tongyi Lab's latest AI video model, and it's a different lineage entirely from ByteDance's Seedance line: a four-mode suite covering text-to-video, image-to-video, reference-to-video, and instruction-based editing, with a built-in planning step and native audio. This guide covers what it is, what it can do, how pricing and access work, and how it stacks up against Wan's own earlier versions and against Seedance.

What is Wan 2.7?

Wan 2.7 is Alibaba Tongyi Lab's flagship AI video model, released in April 2026. Rather than being a single generation mode, it's a suite: text-to-video, image-to-video, reference-to-video, and instruction-based editing are all available through one API, so you can generate a clip, continue it, anchor it to a reference, or revise it without switching tools.

Key features

A four-mode video suite. One model family handles generation, continuation, reference-driven shots, and prompt-based revision, which cuts down on juggling separate tools for each stage of a video workflow. Our AI video API guide has more background on how suites like this compare to single-purpose video models.

Thinking Mode. Before generating, Wan 2.7 interprets and plans your prompt, building a structural understanding of the scene first. That means clear direction on subject, lighting, camera, and mood tends to reward you with more intentional output, rather than the model just matching keywords.

Native synchronized audio. It generates ambient sound, lip-synced dialogue, and background music in the same pass as the video, and it can clone a voice from a supplied audio reference rather than requiring a separate voice model.

First-and-last-frame control. Set a starting frame and an ending frame, and Wan 2.7 generates coherent motion across everything in between, useful for anything where the exact opening and closing shot matters.

Instruction-based editing. Feed it an existing clip and a text instruction (change the background, swap an object, shift the camera), and it applies the edit locally or across the whole clip without re-rendering from scratch. Our AI video editing API guide covers this kind of workflow in more depth.

Multi-reference consistency. It accepts up to five image or video references to hold a character's identity and voice across shots, combined with physics-aware motion and 1080p output, and it runs noticeably faster than previous Wan versions.

Specs at a glance

Wan 2.7 vs Wan 2.5 and 2.6

If you're still on an earlier Wan version, or you've been searching for "Wan 2.5" or "Wan 2.6" and landed here instead, here's what's actually new in 2.7.

Thinking Mode and instruction-based editing are both new to this release. Neither 2.5 nor 2.6 plans the prompt before generating or supports text-driven edits to an existing clip. Multi-reference consistency is also more developed: earlier versions supported fewer simultaneous references, while 2.7 handles up to five image or video references to hold a character's identity and voice across a whole sequence. First-and-last-frame control existed in earlier form but is more reliable here, and generation is meaningfully faster across the board.

If your use case leans on editing existing footage or needs tight character consistency across several shots, 2.7 is worth the move. If you're doing simple one-off text-to-video clips, the practical difference is smaller.

How to use Wan 2.7 (via API)

Using Wan 2.7 through an API comes down to the same pattern as most modern video models: get a key, submit a request, and check for the result.

curl -X POST https://api.apiframe.ai/v2/videos/generate \
  -H "X-API-Key: afk_your_api_key_here" \
  -H "Content-Type: application/json" \
  -d '{
        "prompt": "a cinematic sunrise over a futuristic cityscape, smooth camera push-in",
        "model": "wan-2.7",
        "wan27Params": {
            "resolution": "720p"
        }
    }'

That returns a jobId right away, since generation runs in the background rather than holding the request open (average completion is around three minutes). Check the job status until it shows completed, or register a webhook (an automatic callback to your server when the job finishes) so the result comes to you instead. That's the better pattern once you're past testing.

For image-to-video, pass an image as the first frame. For first-and-last-frame control, add a last_frame alongside an anchor image or clip. For instruction-based editing, point first_clip at an existing video and describe the change in your prompt. Our text-to-video API guide covers the full range of parameters, including negative prompts, prompt expansion, and seeding for reproducible output. For a hands-on walkthrough with real code, see the text-to-video Python tutorial.

Wan 2.7 examples and prompts

A few real prompts that show the range of what Wan 2.7 can do:

Text-to-video: "A 5-second clip of a chef slicing vegetables in a modern kitchen, smooth left-to-right pan, cinematic lighting."

Image-to-video: "A 5-second image-to-video clip animating this product photo, slow turntable rotation, soft studio light."

First-and-last-frame: "A 5-second first-and-last-frame clip morphing a closed flower bud into a full bloom."

Reference-driven consistency: "A 5-second reference-based clip keeping this character consistent as she walks through a busy market."

Native audio: "A 5-second clip of a drone ascending over a mountain river at sunset, orchestral background music."

Instruction-based editing: "Edit this 5-second clip, change the background to a rainy night, and keep the subject and motion intact."

How Wan 2.7 compares to Seedance

Wan 2.7 and ByteDance's Seedance line are two of the most complete video suites available right now, and they solve overlapping problems differently. Wan 2.7's edge is Thinking Mode and instruction-based editing: if your workflow involves revising existing footage through plain text instructions rather than regenerating from scratch, Wan 2.7 is built for that in a way Seedance isn't. Seedance 2.5, on the other hand, pushes further on raw single-pass duration (30 seconds versus Wan's 15) and reference budget (up to 50 versus Wan's 5).

If you're deciding between them, or want the full picture on ByteDance's side, our Seedance 2.0 guide and Seedance 2.5 guide cover that model family in the same depth as this page. If you're comparing across the wider video model field, our roundup of the best AI video generation APIs puts Wan, Seedance, Kling, and others side by side.