MiniMax launches H3, generating 2K video clips up to 15 seconds with native stereo sound

MiniMax/Hailuo video models (H3)

H3 is a rebuilt video model that makes 2K clips of up to 15 seconds with dialogue, effects and ambience generated together with the picture. MiniMax says a redesigned tokenizer and split model workloads let it train faster and price 2K output at under a third of mainstream rivals.

• Accepts text, images, video and audio as references.
• Tokenizer compression gives 4x more effective sequence length.
• Training throughput up nearly 30% from separating understanding and generation work.
• Also outputs 768p at under half the price of rivals' 720p.