Understanding MLOps and LMOps: Evolution, Engineering Practices, and Future Trends for Large Models

The China Academy of Information and Communications Technology (CAICT) began researching MLOps in 2020 and, as large models become mainstream, has shifted focus to LMOps, an operations system tailored for both large and small models.

Key points include: (1) Large models are entering a critical period of scalable application; (2) LMOps is a core engineering element for model deployment; (3) Future trends and outlook.

Large models differ from traditional ML/DL models in scalability, multi‑task adaptability, and plasticity, enabling use cases such as ChatGPT‑style chatbots, autonomous driving, weather prediction, robotics, and multimodal video generation.

The engineering architecture for large‑model deployment is becoming clearer, featuring a multi‑layer service ecosystem (Model‑as‑a‑Service, platform tools, model calling, and application development) that supports development, fine‑tuning, deployment, and AI‑native applications.

Challenges to large‑scale deployment include limited applicability in high‑precision or real‑time scenarios, the need for robust monitoring, maintenance, and cost‑effective resource utilization.

Four pillars are proposed to build a comprehensive LMOps system: technical tools, data governance, operations management, and application development, each addressing data collection, model fine‑tuning, inference acceleration, and resource management.

LMOps expands on traditional MLOps principles—team collaboration, full‑link feedback loops, rapid response, and AI asset management—to support large‑model specific needs such as multimodal data handling, model quantization, and agent frameworks.

Future development will enhance interaction and visualization, close the data‑feedback loop, increase automation and intelligence, optimize resource usage, and expand Agent‑Ops capabilities.

CAICT has already established a complete MLOps standard system, led the creation of LMOps standards, and contributed to international standardization efforts, aiming to provide a clear, high‑quality, and efficient framework for large‑model deployment.