MiniMax's H3Max Generates 5-Second Videos in Under 3 Seconds, Sparking AI Livestream Wave

量子位 · wechat · 2026-08-31

MiniMax partnered with inference acceleration firm fal to launch H3Max, building on the open-source video model H3 (24M downloads in three weeks) with post-training and inference optimization for real-time generation: a 5-second 768p video takes under 3 seconds, with roughly 35x the throughput of H3, while ranking #1 in image-to-video on both ArtificialAnalysis and DesignArena.

Faster-than-playback generation unlocked a new category: within a week, three independent AI livestreams appeared — fal engineer Rehan Sheikh's "transdimensional TV" (5.4M views on X), Pieter Levels' on-demand livestream site, and another fal engineer's interactive continuous-story stream steered by viewer chat commands.

The author's hands-on test: a complex 5-second video took 7 seconds end-to-end via API (including queueing and 1s polling); MiniMax Design generated a 5-scene earbuds ad. H3Max currently caps at 768p with first/last-frame reference only (no subject reference), but finishes clips in 20 seconds that take H3 5 minutes. The piece argues real-time AI video is a watershed for the field, with open-source ecosystem lock-in; Morgan Stanley kept an Overweight rating (target HK$900) and the stock rose 16.18%.

Related event: MiniMax H3 Max: Video Generation Faster Than Playback, Crossing the Real-Time Threshold(23 posts)→

Original post →

More from Multimodal

Multimodal channel →