Overview
- MiniMax publicly released H3 on Friday as a multimodal video-generation model that takes text, images, video and audio and can produce or edit up to 15-second videos at native 2K resolution with stereo sound.
- H3 is available now through MiniMax’s Hailuo AI platform and an API, and the firm said it will publish the full model weights in the coming days under the MiniMax Community License that allows free non-commercial use and limited commercial use for smaller organisations.
- Independent benchmarks reported H3 leads on video-editing tasks but trails Google’s Gemini Omni Flash on text-to-video and falls behind ByteDance’s Seedance and Gemini on some image-to-video tests, showing strengths and gaps that are task specific.
- MiniMax positions H3 for commercial creators in advertising, e-commerce, product design and games and claims 2K generation costs at less than one-third of mainstream rivals while designing the model to run on several Chinese-made chips to reduce reliance on foreign hardware.
- The planned open-weight release extends a growing Chinese open-source movement for large models and could speed third-party innovation, lower costs for creators and shift competitive dynamics for firms and decentralised compute projects that sell access to closed models.