88Z AI, 7 May — ByteDance has open-sourced Mamoda2.5, the world's first 25B-class unified multimodal model. Built on the Qwen3-VL-8B architecture, it achieves state-of-the-art performance in text-to-image, video generation, and editing, with inference speeds significantly surpassing competitors and performance nearing top closed-source models.
At a glance
- World's first 25B-class unified multimodal model open-sourced based on Qwen3-VL-8B
- Inference speed significantly leads competitors, with video editing performance nearing Sora and Kling
The update
ByteDance has officially open-sourced Mamoda2.5, the world's first unified multimodal model with a parameter scale of 25 billion. Built on the Qwen3-VL-8B architecture, it employs a sparse activation mechanism, activating only about 300 million parameters per inference. Mamoda2.5 achieves state-of-the-art performance across all tasks, including text-to-image, text-to-video, and video editing. In terms of efficiency, its inference speed is 12 times faster than Alibaba's Wan2.2 and 18 times faster than Meituan's LongCat Video, with video editing latency as low as 9.2 seconds. Benchmark tests show that the model ranks first in multiple video editing tasks, with overall performance approaching top closed-source models like Sora and Kling.
Sources & editorial note
This article is an AI-assisted synthesis by 88Z AI based on the public sources below. Product claims and availability are attributed to their publishers. It does not represent independent testing or first-hand reporting.
- Mamoda2.5Source published
Source dates above refer to the original announcements; the date under the headline is this article’s publication on 88Z AI.