MiniMax于8月3日正式发布新一代多模态生成模型H3，宣布同步开放模型权重[1]。该模型支持文本、图像、视频、声音的统一理解与生成[1]，最高可生成15秒2K分辨率视频[1]，定价为0.8元人民币每秒[1]。这一举措意在打破视频生成领域的闭源垄断格局[1]。
在性能方面，H3在Artificial Analysis榜单中的视频编辑能力排名全球第一[1]，文生视频和图生视频能力分别位居第二、第三名[1]。MiniMax自研的VAE技术使序列长度相比原有方案减少了四倍[1]。受该消息刺激，港股MINIMAX-W（00100.HK）报收HK$249.4，涨幅超10%[1]。
MiniMax unveiled its next-generation multimodal generative model H3 on August 3, 2026, marking a significant shift toward open-source development in video generation technology [1]. The model integrates unified understanding of text, images, video, and audio, with capabilities to generate videos up to 15 seconds at 2K resolution [1]. Priced at 0.8 yuan per second, H3 ranked first globally in video editing capability on the Artificial Analysis leaderboard as of July 31, 2026, while placing second in text-to-video generation and third in image-to-video capabilities [1].
Concurrent with the launch announcement, MiniMax revealed plans to open the H3 model weights to the public, breaking the closed-source dominance that has characterized the video generation sector [1]. The company's self-developed VAE architecture reduces sequence length by a factor of four, enhancing processing efficiency [1]. MiniMax's Hong Kong-listed stock (MINIMAX-W, ticker 00100.HK) surged more than 10 percent to HK$249.4 following the announcement [1].
