+
+
+
+ One runtime that matches or beats state-of-the-art inference systems. Full methodology and current numbers in the blog post and paper. +
+ ## What is M*? **M\*** (pronounced *"M-star"*) is a serving system for the new generation of **composite multimodal models** — models built from structurally distinct components (vision encoders, transformer backbones, diffusion and flow heads, audio codecs, action generators, world-model predictors) whose execution path changes with the input and the task. diff --git a/assets/mstar_headline.png b/assets/mstar_headline.png new file mode 100644 index 00000000..4e91a6c9 Binary files /dev/null and b/assets/mstar_headline.png differ