Seedance 2.5 on Vercel AI Gateway Brings 30-Second AI Video to Developers
Quick Reads
- ByteDance’s Seedance 2.5 is now live on Vercel AI Gateway
- The model generates up to 30 seconds of video in a single clip
- It accepts text, video, image, and audio as inputs in one request
- Developers can reference images and videos to control motion, appearance, and camera work
- Prompts are supported in over ten languages
- Access is available via the AI SDK using
bytedance/seedance-2.5
ByteDance’s Seedance 2.5 is now available on Vercel AI Gateway, and it changes what developers can expect from AI video generation. The model, announced on August 6, 2026, generates up to 30 seconds of video in a single clip, no stitching, no post-production workarounds.
That matters because most AI video tools still break longer outputs into short segments and join them together. Seedance 2.5 holds camera movement and continuity throughout the clip, producing a result that feels like one seamless shot. Moreover, short clips can be extended while keeping character appearance, scene context, and camera motion intact.
What also sets Seedance 2.5 apart is its multi-input capability. Developers can feed text, video, image, and audio into a single request. In practice, this means pulling a character’s appearance from one asset and borrowing motion or camera work from another, all within the same generation call. Additionally, the model supports local video editing, allowing teams to swap out backgrounds, products, or characters while the frame structure, pacing, and camera angle remain unchanged. One finished cut can therefore produce several versions without re-rendering from scratch.
Prompts work in more than ten languages, which broadens the practical use cases for teams building globally, including markets across Africa where multilingual content workflows are increasingly common.
For developers ready to integrate, the setup is straightforward. Set model to bytedance/seedance-2.5 and call generateVideo from the Vercel AI SDK. Image references are passed under inputReferences and called out in prompts with tags like [Image 1] or [Image 2]. Video references work the same way but are numbered separately, so both [Video 1] and [Image 1] can appear in the same prompt, motion from one source, appearance from another. Vercel’s documentation does caution, however, that mixing reference roles and editing roles in a single prompt produces less stable results. Keeping each asset’s purpose explicit leads to cleaner outputs.
Seedance 2.5 on Vercel AI Gateway is available now. Developers can explore the model page for the full API reference and a live playground. A full list of video models supported on AI Gateway is also available here.
