In-Depth Look at the 2026 Volcano Engine FORCE Original Power Conference: Seedance 2.5 Redefines AI Creation with 30‑Second Native Video

On June 23–24, 2026, the Volcano Engine FORCE Original Power Conference was held at the China National Convention Center in Beijing. As one of the most important annual barometers for China’s AI industry, this year’s event zeroed in on the new trends of production‑grade AI deployment, focusing on generational leaps in large language models, large‑scale commercial adoption of enterprise agents, and the transformation of AI business value. Over the two‑day conference, Volcano Engine unveiled a series of heavyweight models, including the Doubao 2.1 Pro large language model, the Seedance 2.5 video generation model and its 4K enhanced edition, the Seedream 5.0 Pro image generation model, and the Seed‑Audio 1.0 audio generation model. The company also debuted its AI copyright commercialization platform and the Doubao Professional subscription tier. This concentrated flurry of announcements sends a clear signal: domestic large models are crossing the technical “tipping point” faster than expected, evolving from merely “usable” to “robust, durable, and commercially viable.” At the same time, AI video generation has officially entered a new era of long‑form native storytelling and industrialized production.

Liang Rubo Sets the Tone for 2026: Scaling the AI Summit Is ByteDance’s Top Priority

In the opening ceremony, ByteDance CEO Liang Rubo delivered a keynote address via an immersive video, setting the strategic tone for the entire conference. He unequivocally defined ByteDance’s 2026 keyword as “Scaling the Summit” and emphatically stated that climbing the AI mountain is now the single most important and core mission for the company. Liang reviewed ByteDance’s strategic choices over the past few years—continuously narrowing its business scope, concentrating its most elite algorithm teams, ample computing resources, and core R&D budgets on the AI track, and further sharpening its focus within AI to pursue breakthrough improvements in foundational model capabilities. He asserted that the transformative impact of AI will be “at least as significant as the combined changes brought by the PC, Web, and Mobile revolutions.” In this long‑term technological marathon, Volcano Engine’s MaaS (Model as a Service) business has been formally elevated to a foundational core business of ByteDance, backed by a long‑term, steadfast commitment with no preset budget ceiling.

Liang highlighted a key methodology driving ByteDance’s rapid AI iteration: many of the model’s capability enhancements and leaps were not born from isolated lab research, but rather from deep collaboration with clients in real production scenarios. He shared a compelling example—the world‑first “3D white‑box pre‑visualization” feature in the Seedance 2.5 model, unveiled at the conference, originated from a creative spark provided by a leading domestic film director who works closely with Volcano Engine. During pre‑production, the director struggled with the lengthy and costly process of traditional storyboard pre‑visualization and wondered if AI could quickly “preview” the shoot. That real pain point was eventually transformed by Volcano Engine engineers into a groundbreaking feature of Seedance 2.5. This closed‑loop iteration model—rooted in real‑world scenarios and feeding back into them—has become the core engine driving Volcano Engine’s rapid model evolution.

Doubao 2.1 Pro Crosses the “Tipping Point” in Three Major Capabilities with Exceptional Cost‑Effectiveness

As the flagship language model released at this conference, Doubao 2.1 Pro achieved significant generational leaps in three strategic directions: Coding (code generation and understanding), Agent (complex task planning), and VLM (visual language multi‑modal understanding). Honed through massive real‑world business scenarios, the model now firmly ranks among the global top tier across multiple international benchmarks, shattering the previous stereotype that domestic models compete only on price, not performance.

In coding capability, Doubao 2.1 Pro excelled across industry‑recognized challenging benchmarks such as Terminal Bench 2.1, SWE‑Pro, and SciCode, with a SciCode score of 59.8, demonstrating a solid command of complex scientific computing and enterprise‑grade codebases. In agent and multi‑modal tasks, the model ranked among the global leaders on OSWorld, MobileWorld, and MMMU‑Pro, meaning it can not only “understand” images and text, but also operate software, plan itineraries, and orchestrate tools to complete complex, multi‑step tasks like a human. Two live demonstrations were particularly impressive: in one real‑world RTL (Register Transfer Level) chip design verification test, Doubao 2.1 Pro ran stably for nearly 18 hours, completing nine rounds of automated iterative modifications and full design‑rule validation—a testament to its outstanding long‑task stability. In another, a 3D virtual city built on Doubao 2.1 Pro now supports over 500 intelligent agents collaborating simultaneously within the same digital space, offering a high‑fidelity digital testbed for cutting‑edge research in urban management, emergency drills, and economic simulations.

Volcano Engine President Tan Dai stressed that only when a large model crosses the invisible “tipping point” can it meet the stringent demands of enterprise production environments in terms of accuracy, stability, and complex logic. On pricing, Doubao 2.1 Pro continues its cost‑performance leadership: RMB 6 per million input tokens, RMB 30 per million output tokens, and as low as RMB 1.2 for cache‑hit scenarios, bringing total cost down by nearly 80% compared to Claude Opus 4.6. This aggressive pricing will undoubtedly lower the barrier for small and medium‑sized enterprises and independent developers to embrace top‑tier AI capabilities. Doubao 2.1 Pro is now fully open via API and deeply integrated into the Doubao App, the TRAE development platform, and the Coze agent‑building platform.

Seedance 2.5: An Epic Upgrade – 30‑Second Native Video Output Completes the Final Piece for Commercial Creation

If one announcement drew the most attention and the loudest applause at this conference, it was the first public debut of the Seedance 2.5 video generation model. Tan Dai revealed that this eagerly awaited next‑generation model has entered the final sprint of strict internal testing and is expected to officially launch to the public and enterprise users in early July. Its arrival directly rewrites the rules of domestic AI video generation.

Compared to the already impressive Seedance 2.0, Seedance 2.5 delivers three epic innovations that address creators’ core pain points:

First, native video length doubles to 30 seconds. Previously, the native single‑generation length of mainstream domestic AI video models was typically capped at a 10‑ to 15‑second technical bottleneck. To tell a longer story, creators had to split scripts into multiple short clips and then spend considerable effort stitching and color‑grading in post‑production. This traditional path made it extremely difficult to avoid the “AI video syndrome”—facial drift, sudden changes in clothing texture, and inconsistent lighting and camera logic. Seedance 2.5, powered by a new end‑to‑end rendering architecture, now enables smooth generation of a single 30‑second native video with a complete narrative. From the first to the last frame, the model uses a unified physical rule set and rendering pipeline, strictly locking facial features, garment materials, environmental lighting, and camera movement logic throughout the entire 30 seconds. This means a 30‑second brand micro‑film or a complete AI‑generated short story with a clear beginning, middle, and end can now be generated in one go, with no stitching required. At the same time, Seedance 2.5 retains native 4K ultra‑high‑definition output, meeting the stringent quality standards for cinema and premium commercial advertising.

Second, the input capacity for multi‑modal reference materials has been greatly expanded to 50 items. The previous Seedance 2.0 supported at most 12 mixed references (images, videos, and audio), which proved insufficient for producing series‑style short dramas or commercial orders that must strictly adhere to brand visual identity (VI) guidelines. Seedance 2.5 boosts this capacity to 50, fully supporting simultaneous input of images, storyboard sketches, specific reference clips, detailed character design sheets, multi‑angle environment mood boards, specified background music, and ambient sound effects. Creators can now “feed” the model an entire set of character turnarounds and scene concept art at once. Seedance 2.5 then internalizes and unifies all artistic styles, color compositions, and lighting characteristics across the entire generation cycle. This upgrade is especially valuable for maintaining high visual consistency in serialized IP short dramas and batch production of multi‑episode AI animations.

Third, the industry’s first video generation model to feature a 3D white‑box pre‑visualization (blocking pre‑vis) function makes its debut. This is the most disruptive innovation of Seedance 2.5. It allows directors or creators, before committing costly computational resources to final video generation, to quickly generate a low‑cost 3D white‑box “rough cut” that simulates camera movement, character blocking, and keyframe action timing. In traditional filmmaking, this pre‑visualization stage—often called “pre‑vis”—typically requires a professional storyboard team weeks of manual work or expensive 3D software, but Seedance 2.5 compresses it to seconds. Creators can adjust virtual camera angles and experiment with different editing rhythms interactively, just like playing a game, and then trigger the final high‑definition, fully detailed video generation based on the pre‑vis results with one click. As Liang Rubo noted, this feature, inspired directly by the real‑world struggles of a frontline director, was engineering‑realized by the Volcano Engine team and is now empowering the entire creator community, dramatically lowering the cost of trial‑and‑error in film‑grade AI content production.

Additionally, Seedance 2.5 supports more flexible local video editing, allowing users to precisely re‑paint or modify specific areas of a frame (e.g., a character’s accessory, a vehicle license plate, or a specific object texture) without altering the overall composition or atmosphere. Tan Dai emphasized that video generation is not just for entertainment—it is a critical path toward the general world model. Already, the Seedance model family has been deeply deployed in high‑tech industrial sectors such as embodied intelligence robot motion pre‑visualization, digital twin simulation of manufacturing assembly lines, and autonomous driving edge‑case scenario synthesis and data generation, providing a new productivity tool for the digital simulation of the physical world.

A Full Multimodal Matrix Upgrade, Agent‑Era Infrastructure, and the Launch of Doubao Professional

Alongside breakthroughs in language and video models, Volcano Engine also comprehensively upgraded its multimodal model matrix, spanning vision and audio, to build a full‑stack AIGC content production toolkit. The image generation model Seedream 5.0 Pro introduced major interactive upgrades, including interactive precise region editing, automatic multi‑layer separation (greatly facilitating later modifications in Photoshop), and native multi‑language text rendering within images—solving the long‑standing industry headache of AI images failing to generate clear, stable text. The audio model Seed‑Audio 1.0 demonstrated impressive capabilities, not only generating high‑quality background music and foley effects, but also producing complex audio scenes with multi‑character emotional dialogue and environmental sound effects in a single pass, transforming AI videos from “silent movies” into fully immersive “cinematic surround‑sound” experiences.

For the impending agent era, Volcano Engine systematically upgraded its AI‑native cloud architecture, launching the Fangzhou CLI tool, the AgentKit low‑code agent development suite, and the HiAgent 3.0 enterprise platform, along with the enterprise‑grade Agent Workbench and the AI Trust security and compliance framework—delivering end‑to‑end guarantees for enterprises to deploy AI agents at scale, from development and testing to security and compliance. Ecosystem data shared at the conference shows that over 1.1 million enterprises and individual developers are now using the Volcano Fangzhou model service platform, and the number of enterprise clients with annual token consumption exceeding 1 trillion has reached 200, doubling in just six months—a strong testament to the hunger for production‑grade AI capabilities on the enterprise side.

On the monetization front, on the second day of the conference (June 24), Doubao officially launched its Professional edition for power users and productivity scenarios, with a three‑tier pricing strategy precisely targeting different usage intensities: Standard at RMB 68/month, Plus at RMB 200/month, and Premium at RMB 500/month. To nurture future core users, full‑time college students can enjoy a special discount of RMB 38/month for the Standard tier upon verification through the Xuexin academic authentication system. This clear, tiered pricing model marks Doubao’s steady transition from an early‑stage, free‑traffic‑driven general‑purpose entry point toward a mature commercial form driven by high‑value productivity subscriptions.

AI Copyright Commercialization Platform Launches, Stephen Chow’s Classic IP Leads the Way for a New Era of Compliant Derivative Creation

Another milestone with significant industry implications was the public debut and ecosystem signing ceremony of ByteDance’s new AI copyright commercialization platform. Comedy maestro Stephen Chow appeared via video message as one of the “first‑batch partner film creators,” announcing that his vast library of classic film clips would be made available as officially licensed templates on the platform, enabling compliant derivative creation and reinterpretation through Douyin, the Jiemeng AI creation tool, CapCut, and all editing software that integrates with the Seedance ecosystem. This initiative cleverly connects the full loop: “classic copyright library → AI model generation → short‑video distribution → creator monetization.” Tan Dai revealed that since the pilot launch of these classic IP templates, cumulative AI creation calls have already exceeded 100,000—revitalizing dormant classic film assets while equipping the vast UGC creator community with a legal arsenal free from copyright risks.

Stunning Market Data Confirms Leadership as Domestic Large Models Enter Deep Waters

In the closing segment of the conference, Volcano Engine disclosed a series of eye‑catching core business metrics, offering hard numbers to counter lingering skepticism about AI commercialization: as of June 2026, the Doubao model family has achieved a historic daily token call volume exceeding 180 trillion, representing a more than tenfold surge over the past year. In China’s public cloud MaaS segment, Volcano Engine leads with a commanding 49.5% market share. The number of “Trillion Token Club” members—top enterprises across industries—has surpassed 200, covering finance, manufacturing, retail, healthcare, education, and virtually all core sectors of the national economy.

From the early industry question of “whether to do AI” to today’s “how to do AI well,” and from large models that could only “answer questions” to digital productivity tools that can now “complete tasks” independently—the signals from the 2026 Volcano Engine FORCE Original Power Conference are resounding and clear. Artificial intelligence is no longer an abstract concept floating in the cloud; it has become a tangible, measurable, and actionable item on corporate balance sheets for cost reduction and efficiency enhancement. With Doubao 2.1 Pro steadily crossing the production‑grade “tipping point,” and Seedance 2.5 pioneering the completion of the final technical puzzle piece for commercial video creation, domestic large models are delivering on their promise to reshape industries with astonishing engineering speed and implementation resolve. The second half of 2026 marks the starting gun for a fierce race in AI‑driven video industrialization and the large‑scale deployment of agentic workforces—and the competition has only just begun.