Research
MIT Han Lab open-sources Evolving Infra, an agent-adaptive LLM serving system its authors say cuts latency up to 4.1x against NVIDIA Dynamo
Researchers led by Hao Kang and Song Han released Evolving Infra on GitHub on November 10, 2026, alongside arXiv paper 2611.12304. The system lets serving nodes switch between prefill and decode without restarting and us…