Video AI: MiniMax challenges ByteDance with low price, open weights for new H3 model

Chinese AI startup MiniMax has unveiled H3, its latest multimodal video generation model, aiming to challenge the closed-source dominance in the field by offering open weights and competitive pricing. This new model is currently the most powerful AI in video editing, according to benchmark platform Artificial Analysis. However, it falls short of Google's Gemini Omni Flash in text-to-video tasks and behind ByteDance's Seedance 2.0 and Gemini Omni Flash in image-to-video capabilities.

Designed for commercial content creation in film, advertising, video games, and product design, H3 is available through an API and MiniMax's Hailuo AI platform. MiniMax intends to release full model weights soon to support the open-source community and enable users to customize their own versions with various AI hardware. MiniMax's decision to open up H3's weights contrasts with previous proprietary Hailuo video models.

The company released the model's weights under the MiniMax Community License, allowing free non-commercial use and commercial use for organizations with under $20 million in annual revenue, provided proper attribution. This licensing model is expected to become more prevalent as Chinese open-weight AI developers seek to commercialize their massive models.

MiniMax is also targeting aggressive pricing to compete with rival Chinese video AI models. At 768p resolution (1,366 by 768 pixels) with audio, H3 costs $7.80 per minute, while ByteDance's Seedance 2.0 at 1080p (1,920 by 1,080 pixels) costs $22.45 per minute. Despite being more expensive than Google's Gemini Omni Flash, which costs $6 per minute, MiniMax H3 is expected to become the world's strongest open-weight video AI model when its weights become available.

ByteDance also released its new proprietary video model, Seedance 2.5, on the same day, boasting its ability to create 30-second videos in a single run with enhanced editing capabilities. This new model aims to deepen AI use in industries like education and industrial manufacturing, serving corporate clients. ByteDance's annual recurring revenue from large language models has reached $4 billion this month.