Research2026-08-27
The SGLang team published measured benchmarks for running the open-source MiniMax-H3 video model on eight NVIDIA H200 chips. Its exact-match path runs roughly twice as fast as the Diffusers runtime, and combining step reuse with sparse attention goes considerably faster while visibly reducing image similarity to the original output. The results require self-hosting the model on that hardware.
What changed
Running MiniMax-H3 video generation on the Diffusers runtime took 74–217 seconds per clip depending on task and length.
What it unlocks
Generating the same MiniMax-H3 video clips several times faster on the same eight-GPU server, with published settings for choosing between quality and speed.
- 1.85–1.95× lossless vs Diffusers
- up to 6.24× at 0.76–0.91 SSIM
- 8× NVIDIA H200, 50 denoising steps
What you need to act on it
- self-hosted 8× NVIDIA H200 setup
- SGLang Diffusion v0.5.18
- the open-source MiniMax-H3 model
Sources