MiniMax H3 Now Generates Video at Over 2× Real-Time on 8× B200 with SGLang + VDN-H3
Key Info
SGLang-Diffusion with VDN-H3 now enables MiniMax H3 to generate 14.4 seconds of 768p video in just 9.0 seconds end-to-end on 8× B200 GPUs, surpassing 2× real-time speed with no measured quality regression.
Highlights
- 8-step denoising completes in 6.9s on 8× B200, reaching over 2× real-time throughput.
- The 9.0s figure covers the full generation request after warmup, not just the denoising phase.
- No measured quality regression versus the dense 50-step H3 baseline across 103 test prompts.
- Demonstrates how open models compound through open ecosystems, with community optimization directly accelerating MiniMax H3.