Inferact raises $150M to scale vLLM open-source inference engine

vLLM: vLLM is an open-source inference engine originally developed as a side project at UC Berkeley to optimize serving of large language models. It applies virtual memory techniques from operating systems to GPU memory allocation, enabling efficient handling of variable-sized requests. Major technology companies including Google, Amazon, and Meta now deploy it in production environments. Inferact: Inferact is a company founded by the core maintainers of the vLLM open-source project to build a dedicated team advancing inference technology. It focuses on scaling the engine for production use across diverse AI workloads and hardware platforms. The company's work directly stems from addressing ongoing challenges in memory management and scheduling for large language models. Simon Mo: Simon Mo is a researcher and core maintainer of vLLM who transitioned to co-founding Inferact to professionalize development of the engine. He focuses on production challenges such as cache management for agentic and diverse model workloads. Mo remains actively involved in both the open-source project and company operations. Roger Wang: Roger Wang is a researcher involved in creating vLLM and now acts as a core maintainer and co-founder of Inferact. He contributes to technical improvements in the inference engine used by leading AI organizations. Wang helps bridge open-source efforts with enterprise-grade solutions. Kaichao You: Kaichao You is a researcher who helped develop vLLM at UC Berkeley and now serves as a core maintainer and Inferact co-founder. His contributions center on applying operating systems concepts to GPU inference for better efficiency with large models. You supports the project's evolution toward handling increasingly complex AI deployments. Woosuk Kwon: Woosuk Kwon is a researcher who initiated the vLLM project while a PhD student at UC Berkeley to accelerate inference for models like OPT-175B. He continues as a core maintainer and co-founder of Inferact, contributing to ongoing advancements in inference systems. His work emphasizes practical solutions to memory bottlenecks in modern AI serving. Adoption: Leading technology companies including Google, Amazon, and Meta run vLLM in production for their AI workloads. Innovation: vLLM addresses inference challenges from model scale, architectural diversity, and agentic workloads by adapting virtual memory techniques to GPUs. Collaboration: The vLLM project works closely with hardware providers such as NVIDIA, AMD, Google, AWS, and Intel to optimize performance across different accelerators.