SeeDance 2.5: Specs, Release Date, and Everything We Know (2026)

SeeDance 2.5 is ByteDance's next video generation model, announced on stage at the Volcano Engine FORCE conference in Beijing on June 23, 2026. The version number undersells it: one continuous 30-second video per generation, up to 50 multimodal references in a single call, native 4K output with 10-bit color, and a region-editing mode that redraws one element of a finished video without touching the rest. It is not publicly available yet. This post covers what ByteDance announced, when you can expect access, and how to read the claims until independent tests exist.

TL;DR: SeeDance 2.5 was announced June 23, 2026 and is in closed enterprise beta. Public rollout starts in ByteDance's own apps, with third-party API access expected late July 2026. Headline specs: 30-second single-pass videos (vs 15s on SeeDance 2.0), up to 50 mixed references (vs 15), native 4K with 10-bit color (vs 720p/1080p), region editing, and joint audio-video generation. No pricing, variants, or benchmark results have been published. We'll have it on Unifically at API launch; SeeDance 2.0 is live today.

What ByteDance announced

Volcano Engine president Tan Dai introduced SeeDance 2.5 alongside Doubao 2.1 Pro, Seedream 5.0 Pro, and Seed-Audio 1.0. For the video model, the announcement made five concrete claims:

• Native 30-second generation. One pass produces a continuous 30-second video with scene changes and tempo shifts inside it. No stitching. SeeDance 2.0 tops out at 15 seconds per generation, and most competing models sit between 8 and 15.
• Up to 50 multimodal references. Images, videos, audio, style references, and low-fidelity 3D white models in one call. The 3D white-model input is new: a rough blockout can drive motion and spatial layout, which puts previz workflows directly into the generation step.
• Native 4K, 10-bit color. Rendered at 4K rather than upscaled to it, with the bit depth to survive a grading pass.
• Region editing. Redraw one element of a generated video, like clothing, a background, or a product, while performance and lighting stay locked. SeeDance 2.0 has no equivalent; a small change means a full regenerate.
• Joint audio-video generation. Audio is co-generated with the video in a shared latent space, so lip-sync and effects timing come from the model itself. ByteDance lists multilingual output including English, Chinese, Spanish, Japanese, and Korean.

ByteDance also reports roughly 20% better prompt adherence than SeeDance 2.0, and its Dreamina page lists a beta long-video mode that stretches output to 180 seconds.

Release timeline

SeeDance 2.5 skipped the public-launch moment. The rollout is staged:

1. Now: closed enterprise beta.
2. Early-to-mid July 2026: public access through ByteDance's own apps (Dreamina internationally, Jimeng in China), then CapCut.
3. Late July 2026: third-party API access through Volcano Engine, which is when inference platforms and API providers can pick it up.

We will open SeeDance 2.5 on Unifically as soon as the API is live, with the same task-based endpoint used for SeeDance 2.0. The model page is already up and will be updated with callable parameters and pricing on day one.

One wrinkle worth knowing: ByteDance paused SeeDance 2.0's rollout in some regions earlier this year after copyright disputes with major studios, and 2.0 is still not offered in every market. Expect 2.5's rollout to carry the same watermarking and IP guardrails, and possibly the same regional gaps.

How to read the claims

Every number above is ByteDance's own. There are no independent benchmarks yet, no Arena results, and no way to run the model outside the enterprise beta. That doesn't make the claims wrong: SeeDance 2.0's announcement claims held up well once the API opened. But three things should stay open questions until the model is public:

• Generation speed at 30 seconds and 4K. Longer, higher-resolution output usually costs render time. Nothing has been said about latency.
• Pricing. Not announced. SeeDance 2.0 runs $0.11 to $0.13+ per second on Unifically; 4K and 30-second output could price very differently.
• The 50-reference behavior in practice. SeeDance 2.0's 15-reference omni mode is the best multimodal reference surface in production today, and 50 is a different scale. How the model weighs that many inputs is untested.

Arena.ai has no SeeDance 2.5 board entry yet. We will add rankings, head-to-head results, and real output examples to this post and the model page as they land.

What it means if the claims hold

The 30-second ceiling is the big one. Ad spots, product explainers, and short-form story beats mostly live in the 15-to-30-second range, which currently forces stitching: generate segments, cut them together, hide the seams, and hope character consistency survives. A single pass that covers the whole spot with scene changes inside it removes that entire assembly step.

Region editing changes the iteration economics. Today, a client note like "same shot, different jacket" costs a full regenerate and a fresh roll of the dice on everything that was already right. Redrawing one region while the performance stays locked is how editing should work.

And the 3D white-model input is aimed squarely at studios: if a rough blockout can control motion and framing, previz stops being a throwaway artifact and becomes the control surface for the final render.

What to do today

If you're building now, SeeDance 2.0 is live on Unifically in Pro, Mini, and Fast variants: 4 to 15 second videos from 480p up to 4K on Pro, up to 9 image, 3 video, and 3 audio references per call, with native audio. Current rates are on the pricing page. Anything you build against it carries over: 2.5 arrives as a model ID change on the same endpoint, not a new integration.

For the full spec-by-spec comparison, see SeeDance 2.5 vs SeeDance 2.0.