The ByteDance Seed team has introduced Seedance 2.5, a new video creation model released on July 31st. This version is based on the unified multimodal joint audio-video generation architecture of Seedance 2.0 and focuses on improving capabilities in long narrative, multimodal referencing, and editing functions.
The generation duration in single-frame mode has doubled from 15 to 30 seconds, and multi-pass extension support allows for the creation of coherent content lasting several minutes with a unified audiovisual language. The model optimizes transitions between frames and scene changes to ensure the integrity of long videos while enhancing image, sound, and motion quality, which reduces the appearance of oily artifacts often seen in video generation.
Thanks to these improvements, users can produce several-minute content while maintaining a consistent visual style, reducing the need to split frames, re-splice, and process transitions.
Multimodal referencing capabilities have been significantly enhanced. Seedance 2.5 accepts up to 30 images, 10 video clips, and 10 audio snippets as source materials in a single generation. References based on white models, motion, and creative direction are supported. This enables the realization of complex projects involving multiple objects, scenes, and camera angle changes with a deeper understanding of the creative intent.
Furthermore, the model supports precise timestamp control for targeted video editing, strengthening editing capabilities using chroma keying, viewpoint changes, and reference-based editing, which is useful for professional film production and advertising. The architecture integrates multimodal context understanding when creating native dual-channel audiovisual output.
The narrative function represents a shift from simply creating a clip to completing a full creative work. A demonstration example shows a performance in a single take where a singer interacts with staff in a dressing room, passes through a backstage corridor, meets dancers, and goes on stage, after which the camera pulls back, revealing the entire stadium. The model organizes several logically connected frames over 30 seconds, developing from initial setup to progression, climax, and resolution, instead of just extending one scene.
Multi-pass extension preserves character identity, scene environment, and narrative rhythm, ensuring object stability during splices through camera transitions and maintaining audiovisual synchronization throughout the clip.
Seedance 2.5 is becoming available on Jimeng AI and Doubao Pro platforms, with API services coming to Volcano Engine Ark. This release intensifies competition in the AI video generation market against models such as Kuaishou Kling, Alibaba Wan, and international players, positioning ByteDance at the forefront of long AI video creation. The feature set targets both consumers seeking seamless multi-scene generation and professional workflows in film, advertising, and content production that require precise control and reference-based editing. As user expectations shift from creating a single clip to completing an entire creative project, Seedance 2.5 positions long narrative coherence and professional editing control as defining characteristics of the next generation of AI video models.
MiniMax has introduced its first open multimodal generative model, H3, on July 31st, marking a significant event in the field of AI video generation. The model supports 2K audio-video generation with native two-channel sound and a duration of 15 seconds, and it secured the top spot in the global video editing ranking on the Artificial Analysis platform.
The cost for generating videos using the H3 model is set at 0.8 yuan per second, which is only one-third of the prices for comparable flagship products. Following the announcement of the release, the company's shares rose by 14%, recovering previous declines, as investors reacted positively to the cost control strategy.
H3 is capable of understanding contexts related to text, images, video, and audio, while generating audio-video with native two-channel sound. The model's technical advantage lies in its ability to follow instructions, correctly represent textual and brand information, and feature V2V motion transfer—the ability to transfer motion from one video frame to another. This function enables controlled editing of multimodal content.
The H3 model facilitates content generation for various commercial use cases, including advertising, branding, e-commerce, product development, UI/UX, and gaming. This significantly simplifies the process of creating high-quality video content for enterprises and developers. The cost advantage is achieved through systematic optimization: using a highly compressed tokenizer reduces the required number of tokens for video generation, while heterogeneous training configuration, load balancing, and improved GPU utilization substantially lower both training and inference costs while maintaining model quality.
A central element of the release is the open-source strategy. MiniMax plans to make the model weights available within a few days, subject to regulatory compliance. This will allow companies to deploy the model locally and adapt it using their own data to ensure security and compliance. The company stated that chip manufacturers and developers can participate in adaptation and optimization to reduce operational expenses and expand application scope, following the industry trend of moving from closed services to open, collaborative ecosystems.
MiniMax's approach aligns with the example of Moonshot AI's Kimi K3, the largest open-weights model at 2.8 trillion parameters, which reached cluster computing limits within 48 hours of its release and suspended new consumer subscriptions on July 19th for priority use of computational resources by paying users. MiniMax's strategy deliberately avoids the parameter race characteristic of the Chinese AI industry. While leading players face computational limitations and supply shortages, MiniMax focuses on an extreme price-to-performance ratio combined with open-source distribution to build a commercial advantage based on unit economics rather than model scale. Market observers will monitor the community adoption progress of H3 after the open-source release, the pace of commercial orders, and whether the volume-based pricing strategy can return MiniMax to a leading position in the AI race. The model positions economically viable, commercially deployable multimodal generation as a new frontier of competition, and its video editing capability indicates that production quality for professional workflows may become as important as basic generative power.
