MiniMax H3 Hits 2× Real-Time Video Denoising on 8× B200 with SGLang + VDN-H3
Key Info
MiniMax says its H3 video-generation model is getting faster: with SGLang and VDN-H3, it now achieves over 2× real-time denoising on 8 NVIDIA B200 GPUs, generating 14.4 seconds of 768p video in 9.0 seconds end-to-end after warmup.
Highlights
- Claimed throughput exceeds 2× real-time: 14.4s of 768p video is generated in 9.0s after warmup on 8× B200.
- The speedup comes from the SGLang project and VDN-H3 optimizations, with no measured quality regression.
- MiniMax frames the gain as evidence that open models compound through open ecosystems.