MiniMax H3

MiniMax H3 is a general-purpose omni-modal generation model released by MiniMax on July 31, 2026, and open-sourced days later at just 33B parameters — an uncommon pairing of SOTA-level video performance with open weights . Rather than treating text-to-video, motion transfer, and audio-video editing as separate tasks, it unifies understanding across text, image, video, and sound in one context, outputting up to 15 seconds of 2K video with native stereo audio, aimed at commercial scenarios like advertising, e-commerce, gaming, and product design . In blind tests run by 葬AI, H3 unexpectedly took first place, scoring close to Seedance 2.0 and holding a 55% win rate over Seedance 2.5 ; per 葬AI's later piece, it became "the only model that can fight outside of Seedance," overtaking Keling and pulling MiniMax back from the brink, while contributing the company's largest revenue share in August . Traction has been real: per Vital Capital (绿洲资本), three weeks after open-sourcing it passed 24 million downloads with 300+ public derivative models, and its faster sibling H3 Max has pushed into real-time settings like Twitch streams and 24-hour AI TV channels .