M*: New Serving System for Multimodal AI Models Shows Performance Improvements
Researchers have developed M*, a universal serving system designed to efficiently handle composite AI models that integrate multiple components like vision encoders, language models, and diffusion systems. The system represents models as dataflow graphs and uses a modular abstraction called the Walk Graph to optimize performance across diverse model architectures. M* demonstrates significant efficiency gains—20% lower latency for text-to-image tasks and up to 2.9x improvements in throughput for text-to-speech workloads compared to existing frameworks.
M* addresses a gap in existing model serving infrastructure by providing a universal framework for the new generation of composite multimodal models that combine diverse components such as vision encoders, language backbones, diffusion heads, audio codecs, and world-model predictors. The system represents these complex architectures as dataflow graphs, allowing requests spanning multiple modalities to be processed as traversals over these graphs. The core innovation is the Walk Graph abstraction, which enables modular composition of model components, flexible placement across distributed clusters, and model-agnostic optimizations within a distributed runtime. Benchmarks show M* achieves 20% lower end-to-end latency than vLLM-Omni for text-to-image workloads on BAGEL, delivers up to 2.9x lower real-time factor and 2.7x higher throughput for text-to-speech on Qwen3-Omni, and outperforms V-JEPA 2-AC baselines for robotic planning by up to 12.5x. The work aims to reduce developer effort in deploying increasingly complex AI systems.
What's missing
The paper does not discuss potential limitations of the Walk Graph abstraction for certain model architectures, computational overhead of the dataflow graph representation, or how M* handles dynamic model composition and real-time model updates. Additionally, the scope of evaluation is limited to specific model families and workloads, and generalization to other multimodal architectures is not addressed.
What different sources said
- arXiv cs.AICenter
M*: A Modular, Extensible, Serving System for Multimodal Models
Related
Topology-Aware Thermodynamics Improves DNA Probe Specificity Design
Researchers developed a new framework for designing DNA probes that accounts for the spatial organization of matched sequences, not just overall thermodynamic stability. Traditional methods rely on scalar measures like melting temperature and free energy, which miss how mismatches are distributed along the probe. The approach could improve diagnostic accuracy in applications like HPV detection and gene expression profiling.
Study Identifies Optimal Thermal Dose for Combining Focused Ultrasound with Immunotherapy in Tumors
Researchers used multimodal PET imaging to identify an optimal thermal dose range for focused ultrasound ablation that destroys tumor tissue while preserving conditions for immunotherapy delivery. The study found that excessive heating collapses blood vessels needed for antibody access, while insufficient heating fails to adequately reduce tumor burden. The findings could guide clinical design of combination treatments pairing thermal ablation with immunotherapies.
Plant MSH1 Protein Functions as Mismatch-Directed Nuclease for Organelle Genome Maintenance
Researchers have identified the precise mechanism by which the AtMSH1 protein in Arabidopsis plants recognizes and cleaves DNA mismatches and lesions, preventing mutations in organellar genomes. The protein combines a DNA mismatch recognition module with a nuclease domain that makes staggered cuts at specific positions relative to DNA damage. This discovery explains how plants maintain unusually low mutation rates in their mitochondrial and chloroplast DNA compared to other eukaryotes.