What Is Seedance 2.5? ByteDance's Multimodal Video Model, Explained

Seedance 2.5 is ByteDance’s next-generation video model. It is designed to take a mix of text, images, video clips and audio files as input, and return one continuous video with sound generated in the same pass.

Access today is pre-release and provider-mediated — the Seedance 2.5 API is not generally available. The capabilities below reflect what is expected at launch; treat anything unconfirmed as pending and verify it against the provider’s own documentation before you rely on it in production.

The headline changes (expected)

1. Longer single takes. Seedance 2.5 is expected to support durations of roughly 4 to 30 seconds in a single continuous take, reducing the need to stitch clips and fight continuity breaks — characters drifting off-model, lighting jumping between cuts. Confirm the exact supported range with the provider before building around it.

2. Launch resolutions: 480p and 720p. At launch the model is expected to output 480p and 720p. Higher resolutions (1080p and above) are pending until provider documentation confirms them — do not assume anything beyond 720p is available yet.

3. Multimodal references. You can condition a generation on multiple reference files: up to 30 images, 10 videos and 10 audio references per request. In practice that means you can pin characters with reference photos, lock a color palette, echo a camera move from a clip you upload, and guide pacing with a music track — all in one request. You stop describing a video and start assembling its ingredients.

Audio is generated with the picture, not after it

Seedance 2.5 is designed to produce video and sound in a single pass, and native audio is included at no extra charge. Dialogue, foley and score arrive synchronized with the motion — and if you provide an audio reference, movement and pacing can follow it. Audio generation is controlled by a generate_audio flag (on by default).

What it costs

No public rate card has been published yet. What we can say about the expected model:

• Billing is per second of output.
• Native audio is free — no surcharge for the synchronized soundtrack.
• Turning content_filter off adds roughly 10% to the cost of a generation.

Public rates are expected to launch alongside the API, so verify the live pricing in the provider console before committing to a budget. Our pricing page tracks what is known, and the playground lets you assemble a request before you generate.

How to access it

Access is pre-release and provider-mediated. The practical route for developers and AI agents is a third-party API provider such as EvoLink, which is expected to serve Seedance 2.5 over a REST API. Confirm provider support and availability before you build on it. The API quickstart has working code, and the agent guide covers automated workflows.

The API contract (pre-release — keep it configurable)

Because this is pre-release, treat the model id as configurable and re-check every field against the provider docs. The expected contract:

Submit a generation with POST https://api.evolink.ai/v1/videos/generations:

curl -X POST https://api.evolink.ai/v1/videos/generations \
  -H "Authorization: Bearer $EVOLINK_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "seedance-2.5-reference-to-video",
    "prompt": "A cinematic drone shot pulling back over a coastal town at golden hour",
    "duration": 5,
    "quality": "720p",
    "generate_audio": true,
    "content_filter": true,
    "image_urls": [],
    "video_urls": [],
    "audio_urls": [],
    "callback_url": "https://example.com/webhook"
  }'

Field reference:

• model — seedance-2.5-reference-to-video (pre-release; keep it configurable).
• prompt — your text direction.
• duration — 4 to 30 seconds (default 5).
• quality — "480p" or "720p" (default "720p").
• generate_audio — boolean, default true.
• content_filter — boolean, default true; setting it false applies a 1.1x cost multiplier.
• image_urls — up to 30 image references.
• video_urls — up to 10 video references.
• audio_urls — up to 10 audio references.
• callback_url — optional webhook for completion.

Then poll for the result with GET https://api.evolink.ai/v1/task-detail?id=<id>. Status moves through pending → processing → completed (or failed).

Where it stands against the field

The Seedance family has been competitive in public video-generation benchmarks, and 2.5 is expected to extend that with longer single takes, native audio and a larger multimodal reference budget. Against models like Sora 2 and Veo 3.1, the practical differentiators are expected to be the multi-reference input and single-take coherence — but keep in mind these capabilities are pre-release. For prompt structure that works well with the model, see the prompt library.

The short version

Seedance 2.5 aims to be a video model you can “direct like a shoot”: you bring the cast (image refs), the camera move (video refs), the score (audio refs) and the direction (prompt), and it returns one coherent, sound-synced take. It is pre-release and provider-mediated today — expected durations of 4 to 30 seconds, launch resolutions of 480p and 720p, up to 30 image / 10 video / 10 audio references, native audio at no extra charge, and per-second billing with rates to be confirmed at launch. This is an independent guide; EvoLink is the recommended third-party provider. Verify every detail with the provider before you ship. To see how a request comes together, start at the playground.