A team of researchers has built a serving system that treats multimodal AI models as dataflow graphs, claiming meaningful speed gains over vLLM on several benchmarks.
As AI models have grown to combine vision, language, audio, and robotics components into a single architecture, the frameworks built to serve them have not kept pace. M addresses this by representing models as "Walk Graphs," a directed dataflow abstraction that lets individual components be placed flexibly across a compute cluster and optimized independently. In tests, M achieved 20% lower end-to-end latency than vLLM-Omni for text-to-image workloads on BAGEL, delivered 2.9x lower real-time factor and 2.7x higher throughput for text-to-speech on Qwen3-Omni, and outperformed a V-JEPA 2-AC baseline for robotic planning by up to 12.5x.
vLLM has become the default serving layer for most production LLM deployments, but it was designed around single-modality language models. The rise of so-called omni models, systems that process text, image, audio, and action outputs in a single pipeline, has created a genuine mismatch: the infrastructure still assumes a simpler architecture than what actually ships.
The paper is research, not a product you can pull from a registry today, and benchmark numbers from authors comparing their own system to a well-known competitor deserve the usual skepticism. That said, the underlying diagnosis is hard to argue with: serving infrastructure is quietly becoming a bottleneck as model architectures add components faster than the tooling can adapt.