Show HN: ElasticMM – 4.2× Faster Multimodal LLM Serving (NeurIPS 2025 Oral) (github.com)

ElasticMM, a novel serving system for large multimodal models (LMMs), was introduced, achieving up to 4.2× lower time-to-first-token (TTFT) and 3.2–4.5× higher throughput compared to existing frameworks like vLLM. This innovative system incorporates Elastic Multimodal Parallelism (EMP), allowing for dynamic GPU allocation between text and multimodal workloads, real-time auto-scaling, and a two-level hierarchical scheduling architecture for optimal resource management. Developed with a focus on efficiency and scalability, ElasticMM supports multiple GPUs and facilitates simple integration with applications through an OpenAI-compatible API. The significance of ElasticMM for the AI/ML community lies in its potential to enhance the processing capabilities of multimodal models, which are increasingly important for advanced AI applications. By optimizing resource utilization and reducing latency, ElasticMM addresses key challenges in deploying LMMs. The open-source project is built on the vLLM foundation and offers valuable extensibility through a calibration process designed to tailor performance to specific hardware configurations. Researchers are encouraged to utilize and cite the ElasticMM framework, contributing to a more efficient and accessible landscape for multimodal AI advancements.