M* Runtime Launched to Unify Composite Multimodal Models

jure · x · 2026-07-06

The newly introduced inference runtime M argues that modern multimodal models are no longer just single decode loops, but composite systems. M provides a unified runtime to serve various multimodal models. The creators claim it achieves roughly 2.7× speedup on omni TTS and 12.5× on world model rollouts, matching or beating specialized systems.

Original post →

More from Infra

Infra channel →