🧠 Food for thought

Implications, context, and why it matters.

The ‘Mini’ in MiniCPM shapes its strategy

• ModelBest started the MiniCPM family as small language models (SLMs) with 2.4 billion and 1.2 billion non-embedding parameters, leaning on efficiency rather than sheer size 2.
• That choice lets some workloads run on devices, which can cut cloud-based application programming interface (API) calls for certain tasks and reduce ongoing inference costs 3.
• The lineup now includes 8 billion-parameter multimodal models with features such as real-time video understanding, while the guiding idea stays centered on efficient local execution 45.
• Open-source work with Tsinghua University’s THUNLP, a natural language processing lab, and Renmin University of China helps explain download momentum on GitHub and Hugging Face 6.

On-device AI offers another path alongside cloud systems

• ModelBest has put MiniCPM on-device models into automobiles, smartphones, personal computers (PCs), and smart home devices, which can sit alongside cloud-first setups in some deployments 7.
• In autos, ModelBest has named Geely, Volkswagen, Changan, and GAC Group, a Chinese state-backed automaker, as partners where low latency and spotty connectivity make local AI practical 8.
• ModelBest says MiniCPM is built into Changan Mazda EZ-60 and Geely Galaxy M9 1.
• ModelBest positions MiniCPM-o as a move toward “full-modality” systems, meaning AI that handles multiple input types. The company says it launched MiniCPM-o 4.5 and presented MiniCPM-V 4.5 on a robot at the 2026 Zhongguancun Forum, a major Chinese technology conference 91.