ByteDance confirms Seedance 2.5: 30 seconds of native video, 50 references and release in July

ByteDance has officially confirmed the name of its upcoming AI video model: Seedance 2.5 (Doubao-Seedance-2.5). The announcement came on June 23, 2026 during the FORCE summer conference of Volcano Engine—ByteDance's cloud division—in Beijing, where president Tan Dai (谭待) presented the model as the evolution of the Doubao family aimed at professional production. According to Sina Finance and media citing 每日经济新闻, Seedance 2.5 is in global enterprise beta and should reach production environments in early July.

What ByteDance confirms with Seedance 2.5

Until now, rumors and leaks pointed to a mid-term review of Seedance with more duration and resolution. The FORCE conference removes the ambiguity: the product is called Seedance 2.5 and it is not a minor patch. Volcano Engine positions it as a leap from experimentation with short clips to a production-ready flow in film, advertising, e-commerce and industry.

The slide shown at the event—collected by attendees and media such as AIBase— lists three structural improvements:

• 30 s of native video in a single segment ("单段原生视频直出"): the model generates up to half a minute without concatenating multiple renders, a ceiling that most competitors still do not reach natively.
• Up to 50 multimodal references: images, videos, audio and style layers in a single generation, compared to the approximate limit of 12 assets in Seedance 2.0.
• Local editing with visual consistency: modify characters, clothing or objects within a scene without regenerating the entire clip or losing lighting, physics or composition.

Seedance 2.5 实现了30秒单段原生视频直出、最多支持50个全模态素材 联合生成、保持画面一致性的局部编辑三项能力提升。

— Sina Finance, June 23, 2026

Seedance 2.5 vs. Seedance 2.0

Media such as Lapaas Voice and TMTPOST add technical details not contradicted by Chinese sources: director-level camera control, phoneme-level lip sync, automatic ambient sound design, and a beefed-up physics simulation engine for long multi-shot sequences.

FORCE 2026: beyond the name

The conference did not limit itself to naming Seedance 2.5. Volcano Engine took advantage of the scenario to reinforce its multimodal and business commitment:

• Seedream 5.0 Pro: image model with interactive editing, layer separation and native multilingual text.
• Seed-Audio 1.0: Generative audio with multimodal zero-shot reference, multi-role dialogue, music and foley effects.
• Doubao 2.1 (Pro and Turbo): coding and agent-oriented models, with prices from 3–6 yuan per million input tokens.
• AI Copyright Platform: ByteDance anticipated a market to monetize officially licensed IP; PANews cites 周星驰 (Stephen Chow) as the first creative partner, with templates authorized in Douyin, Jimeng and CapCut.

Tan Dai stressed that video generation is "one of the paths to world models" and that Seedance is already used in robotics, manufacturing and autonomous driving for data synthesis, scenario simulation and process demonstrations, according to Sina Finance.

Volcano Ark—the cloud model service—now exceeds 1.1 million business and personal users, with 200 companies exceeding one billion tokens per year, double what it was six months ago, according to the same conference.

What it means for CapCut, Dreamina and creators

Seedance 2.0 is already integrated into CapCut, Dreamina and Volcano Ark. The confirmation of Seedance 2.5 anticipates a new wave on those same platforms, probably first in Chinese enterprise accounts and then in the international rollout that ByteDance usually does in phases — as happened with Mini in June.

For a TikTok or Reels creator, the jump from 15 to 30 native seconds eliminates one of the biggest frictions: assembling multiple AI clips and hiding cuts. For agencies, the 50 references allow you to anchor brand identity, character and visual style in longer narratives without drift. And local editing reduces the cost of iterating when only one detail needs to be changed.

The competition—Sora, I Spy, Kling, Runway—presses on cinematographic quality; ByteDance responds with distribution (CapCut exceeds 200 million monthly users) and now with duration and control in a single already confirmed product name.

In summary

Is it confirmed? Yes: ByteDance and Volcano Engine officially use the name Seedance 2.5 / Doubao-Seedance-2.5. When? Enterprise beta now; Wide release planned for early July 2026. What does it provide? 30 native seconds, 50 references, local editing and (according to reports) 4K. And Seedance 2.0? It is still active and gains native 4K generation in parallel. Official sources: Volcano Engine · Seed by ByteDance.