Skip to main content

From Typewriter to Smart Toy: The Edge AI Revolution Led by a 28-Year-Old CTO

The artificial intelligence wave is crashing hard, but while most players are locked in a cloud computing arms race, one startup is zigging while others zag. Mianbi Intelligence has bet big on a counterintuitive idea: shrink those massive AI models down so they can run smoothly on your phone, your car, or even a cuddly toy.

At the helm of this technical revolution is Zeng Guoyang, the company's 28-year-old co-founder and CTO. His AI journey started early—really early. Back when he was just 22, he led the team that trained CPM-1, China's very first large language model. Back then, a simple webpage—jokingly called "a typewriter"—gave the first wave of AI researchers a glimpse of what generative models could do. Over the next few years, he watched the architecture evolve from BERT to GPT, and became convinced that generative AI was the path to something bigger.

Image

Today at Mianbi Intelligence, Zeng and his team are obsessed with what they call "knowledge density." Their core belief: simply piling on more parameters isn't the only way forward. Through their "model wind tunnel" technology, they can efficiently test and predict model performance in small-scale experiments. The math behind it is striking—knowledge density doubles every 3.5 months, meaning the parameter count needed for the same level of intelligence drops exponentially. Take their MiniCPM model: with just 2 billion parameters, it outperformed its 8-billion-parameter rivals at launch, carving out a solid niche in the edge computing market.

The logic of AI deployment is shifting, Zeng argues, from "cloud computing power" to "deep understanding." Edge models don't just need to solve engineering headaches like power consumption, latency, and hardware compatibility—they also need genuine, personalized memory. He talks about the concept of the "默契 system" (intuitive system): future AI shouldn't just mechanically respond to commands. Instead, it should adjust your room temperature or plan your travel route before you even say a word. That kind of "invisible" intelligence, he believes, is the ultimate form of edge AI.

To get there, the team is rethinking the entire training process from the ground up. They've built a training framework called ForgeTrain and established a five-tier hierarchy that covers everything from data governance to hardware deployment. Zeng emphasizes that data quality sets the ceiling for model performance—every algorithm engineer has to dive deep into the data layer to make sure the knowledge fed into the model is spotless.

Key Points

  • Mianbi Intelligence focuses on compressing large models for edge devices like smartphones, cars, and toys.
  • CTO Zeng Guoyang led China's first large language model at age 22 and now champions "knowledge density" over raw parameter count.
  • Their MiniCPM model (2B parameters) outperformed 8B-parameter competitors at the same time.
  • The company's "默契 system" envisions AI that anticipates user needs without explicit commands.
  • ForgeTrain framework and five-tier data hierarchy ensure high-quality training data and efficient model deployment.