Tencent's AI just broke a famous conjecture by math legend Yau

Tencent Hunyuan, the Chinese giant's AI division, released Hyra 2.0 on November 2, its research agent for science. The big change is that the agent no longer just solves problems one at a time. It designs and retrains scientific models and tools that can then be reused again and again.

The loudest result comes from math. Shing-Tung Yau, one of the world's best known mathematicians and a Fields Medal winner, has long been credited with the conjecture that every closed manifold with positive curvature admits a smooth circle action, a hidden rotational symmetry. Hyra's HyProver pipeline found a counterexample: a seven-dimensional positively curved manifold with no such action. Tencent says mathematicians checked the proof, and the paper is public. One caveat: the manifold is not simply connected, so the stricter version of the conjecture still stands.

There is a second result too. Petty's conjecture on equidistant points had only been settled up to dimension 4, back in 2005. Hyra pushed it to dimension 5: every five-dimensional normed space contains six points that are all the same distance apart.

Then came the practical stuff. The agent trained HyInv from scratch, a protein design model with just 1.96 million parameters that, by Tencent's tests, beats ProteinMPNN and nine other known models on quality and runs dozens of times faster than the reference sampler. Weights and code are on GitHub, along with HyPLM-Flash, a 650 million parameter protein language model. The quantum tool HyQEC built an error correcting code about 48% more efficient than the previous leader of the qLDPC Challenge. And China's XiangShan RISC-V processor got 5.79% more performance per gigahertz after the agent's tuning.

Tencent says Hyra is already helping build the next Hy model. An AI that trains the next AI and settles questions people have chewed on for decades is no longer a slide in a keynote.