The wave of artificial intelligence is surging forward. While most forces are racing in the cloud computing arms race, Mianbi Intelligence has chosen a completely different path: compressing large models so they can run smoothly in smartphones, cars, and even a plush toy. As co-founder and CTO of Mianbi Intelligence, 28-year-old Zeng Guoyang is personally experiencing and driving this transformation in edge AI.

Zeng Guoyang's journey in AI began very early. At the age of 22, he led the training of China's first large language model, CPM-1. At that time, a simple web page humorously called "a typewriter" allowed the first group of AI researchers to foresee the power of generative models. In the following years, he witnessed the architectural evolution from BERT to GPT, and firmly believed that generative AI would be the path to higher intelligence.

image.png

Today, at Mianbi Intelligence, Zeng Guoyang leads his team to focus on "knowledge density." They believe that simply increasing parameters is not the only way for AI development. Through their "model wind tunnel" technology, they have efficiently validated and predicted model performance in small-scale experiments. The core of this methodology is: knowledge density doubles every 3.5 months, and the parameter scale required for the same level of intelligence decreases exponentially. Take Mianbi's MiniCPM as an example, which, with only 2B parameters, outperformed its 8B competitors at the same time, successfully securing a place in the edge market.

The core logic of AI deployment is shifting from "cloud computing power" to "deep understanding." In Zeng Guoyang's view, edge-side models must not only solve engineering challenges such as power consumption, latency, and hardware compatibility but also possess genuine personalized memory. He mentioned the concept of the "默契 system" (intuitive system): future AI should not just mechanically respond to commands, but instead adjust the room temperature or plan travel routes before the user even speaks, and this kind of "invisible" intelligence is the ultimate form of edge AI.

To achieve this goal, the team is deeply rethinking the training process. Mianbi Intelligence has developed a training framework called ForgeTrain and established a five-tier hierarchical standard from data governance to hardware deployment. Zeng Guoyang emphasized that data quality determines the upper limit of the model, and every algorithm engineer must go deep into the data layer to ensure that the knowledge input into the model is flawless.