MiniMax

AI Video Generation · Generative AI · Open-Source AI

Launch

MiniMax launches H3 video model with open weights, native audio

It arrives days after ByteDance's Seedance 2.5, intensifying a price war among Chinese video-generation labs racing for developer adoption.

• MiniMax released Hailuo 3.0, internally called MiniMax H3, an open-weights video generation model that produces clips up to 15 seconds at up to 2K resolution with native stereo audio, launched July 31, 2026.
• MiniMax is a Chinese AI lab behind the Hailuo video model line, which the company says has already generated more than 590 million videos across its prior two generations.
• H3 accepts text, images, video and audio as reference inputs, letting it perform generative video editing—reworking an existing clip instead of only generating new footage from scratch.
• The model outputs at 24 frames per second with 32kHz stereo audio, supports 11 languages, and accepts up to 9 reference images or 3 video clips per generation, per MiniMax's spec sheet.
• Benchmark platform Artificial Analysis ranked H3 the world's most powerful model for video editing, though it trails Google's Gemini Omni Flash in text-to-video and lags behind Gemini and ByteDance's Seedance 2.0 in image-to-video tasks, according to the South China Morning Post.
• By open-sourcing weights alongside aggressive pricing, MiniMax is betting that developer access and customization—not just benchmark scores—will decide which lab wins China's video-generation race against ByteDance and Google.