The growth dividend from static internet data is topping out. YuanKong Intelligence offers a judgment: the next leap in model capability will increasingly depend on the “edge” that continuously interacts with real environments. Spun out of Peking University, the company gained attention three years ago with ChatExcel, one of China’s earliest AI Excel products. It has since expanded from cloud applications to local, on-device AI.
Recently, YuanKong launched the Boxer on-device model, YuanKong AI Work for office scenarios, and YuanKong AI Science, a research agent, creating a closed loop across models, agents, and devices. A more direct market signal: its on-device model has been selected for preinstallation on HP business computers, taking the university-lab-born team into devices from a leading global hardware manufacturer.
The team redefines on-device AI as “productivity = YuanKong edge AI = model × agent × device.” That means edge AI is no longer a one-way pipeline for “moving models down” to devices; it is a flywheel that can turn on its own: devices generate real interaction data, agents complete tasks locally and produce feedback, and models update continuously based on that feedback. Rather than simply cramming stronger models into devices, this approach emphasizes evolution through long-term operation.
The rationale rests on data-side pressure. Stanford’s AI Index 2026 report raises the idea of “Peak Data,” pointing to high-quality human-generated data approaching the point of being fully consumed by frontier models. Research from Hugging Face and other teams has also simulated performance decay