MiniMax H3Max generates 5s 768p video in under 3s, enabling real-time AI streaming

MiniMax 稀宇科技 · wechat · 2026-08-31

Real-time generation threshold reached

MiniMax has brought H3Max 768P/480P to its open platform: a 5-second 768p audio-video clip takes under 3 seconds to generate, 15s clips finish in 15s, throughput is 35x that of H3, and it ranks #1 on Artificial Analysis and DesignArena image-to-video leaderboards. Generation now runs faster than playback, leaving headroom for queueing, moderation and streaming.

Ecosystem experiments

FastH3 community optimization

FastVideo, NuvaLab and NVIDIA FastGen released FastH3, compressing H3's 49 transformer forwards to 4 with 90%-sparsity VSA: 8x B200 generates a 15s 768p clip in 13s, up to 14x speedup on a single Blackwell GPU; weights, LoRA and inference recipes are public.

Open-source traction

H3 has topped 24M downloads and 300+ public derivatives in three weeks, which MiniMax calls the most downloaded model of 2026. fal added new data and verifiable RL on top of the open weights for prompt adherence and real-time inference.

Original post →

More from Infra

Infra channel →