SGLang

Moore Threads announced that its MTT S5000 AI card and MUSA software stack were adapted for MiniMax H3-Image on the same day the model was open-sourced, enabling single-card generation of 4096×4096 high-resolution images. The adaptation reuses SGLang, MATE and MCCL components and targets the heavy attention overhead of diffusion models under multi-reference inputs, laying groundwork for Day-0 support of future image and video generation models.

Flide is a high-performance exact inference engine for Jev-style decision models, claiming up to 9.5x speedup over vLLM and SGLang via exact state forking while preserving single-question-equivalent outputs. It supports 30+ local decision models, exposes a /v1/systemone API, and installs via pip.