MiniMax’s commanding 82% implied probability stems from its H3 model’s top rankings on recent image-to-video leaderboards, including strong Elo scores on Artificial Analysis and LMArena for motion quality, character consistency, and native stereo audio generation. The 33B-parameter multimodal system, released in July and open-sourced shortly after, supports up to nine reference images, first-and-last-frame conditioning, and synchronized 11-language dialogue in clips reaching 2K resolution. Optimizations such as faster inference variants and community fine-tunes released in mid-October have reinforced trader confidence in its edge over Google’s Veo 3.1 and lower-probability alternatives like ByteDance’s Seedance or Alibaba’s Wan models. With resolution imminent, recent benchmark dominance and reference-handling features remain the decisive factors.