Intel / post

M* Runtime Unifies Serving for Composite Multimodal Models

Source: x.com
Date: Published Jun 18, 2026

Context

Modern AI systems increasingly chain multiple components together, such as a speech-, an omni model, or a vision-language-action model, rather than running one -by-token decode loop the way a standard LLM server assumes. Stanford SAIL's post points to M*, which instead represents a request as a traversal over a dataflow graph, letting one runtime serve these composite pipelines.

The authors report M* reaching up to 2.9x lower real-time factor and 2.7x higher throughput for text-to-speech workloads on Qwen3-Omni compared with vLLM-Omni, and up to 12.5x faster robotic-planning rollouts compared with a V-JEPA 2-AC baseline. These are the paper's own comparisons against specific named baseline systems, not a claim that M* outperforms every possible serving approach.

Terms in this piece · Glossary

multimodal — A model that works with more than text — reading images, audio, or video, and sometimes generating them too.
LLM — A large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.
token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
inference — Running a trained model to get answers — the phase where AI is actually used, as opposed to trained.

Read the source x.com