MiniMax's newly unveiled Hailuo 02, a second-generation video AI model, is reported to exceed Google's Veo 3 in user benchmarks while significantly reducing video costs. The model, utilizing a noise-aware compute redistribution (NCR) architecture, showcases enhanced performance and efficiency.

MiniMax has launched Hailuo 02, boasting three times more parameters and four times more training data than its predecessor. The NCR architecture, a key component of Hailuo 02, optimizes training by compressing videos during early stages with artificial noise and processing them at full resolution later.

MiniMax claims that Hailuo 02 excels in handling complex prompts and accurately simulating physical processes, asserting it as the only model capable of generating intricate scenes such as gymnastics routines.

While MiniMax highlights the NCR architecture's importance, detailed technical specifications have yet to be disclosed. The advancements in Hailuo 02 are attributed to both architectural improvements and enhanced data quality and diversity.

Experts anticipate that Hailuo 02 could potentially democratize high-quality video generation, making it more accessible due to its lower costs. The competition between MiniMax and Google signals a growing trend towards more efficient and performant AI video models.