Three months ago, Tencent Hunyuan launched Hyra1.0, a minimalist agent loop system for general scientific research problems. Powered by its open-source model Hy3, Hyra successfully solved a core problem in additive combinatorics that had remained open for more than fifty years—the optimal exponents of sum sets and difference sets—setting a historic record as the first Chinese open-source model to fully automatically solve an open problem in mathematics. Soon afterward, international AI giants also began competing in mathematics and theoretical computer science.

Now that the path of “AI problem-solving” has been validated, where should automated scientific research go next? How can AI solve complex scientific problems more economically and sustainably? Facing these industry questions, Tencent Hunyuan’s answer is: upgrade the tools. AI4Science has evolved from problem-specific systems such as AlphaFold, AlphaEarth, and AlphaGeometry to a paradigm in which general models drive agents to directly solve scientific problems. The newly released Hyra2.0 aims to go one step further, letting general agents deeply participate in the R&D and upgrading of scientific tools and models, so that the results of one R&D effort can be repeatedly used in subsequent research, creating intelligent compounding.

image.png

A New Harness Architecture Built for Upgrading Tools