The system functions as a three-layer architecture of overall, mission, and worker agents, which are coordinated through a persistent task queue and a shared research board.
Code and research results are published in an Apache 2.0 GitHub repository. Tencent says Hyra is now helping train its next Hy models. Tencent claims HyInv runs 57 times faster than ProteinMPNN while achieving superior structural agreement.
The HyPLM-Flash protein language model, which contains 650 million parameters and was trained on 726 billion tokens, processed approximately 130,000 residues per second on a single H20. Among 12 models compared, HyPLM-Flash was the fastest and ranked third on average across 35 tasks. Tencent says its Largo model achieves the lowest energy error at 0.364 meV per atom.
The suite successfully simulated an aquaporin system containing 60,000 atoms for a duration of 100 nanoseconds. HyProver settled Petty's equilateral-set conjecture in dimension five, increasing the highest settled dimension from four. A GitHub write-up for the agent describes a counterexample to the unrestricted form of Yau's circle-action conjecture, identifying a closed seven-manifold that is not simply connected.
HyQEC identified a [[488,123,22]] code. Tencent claims this code beats the previous leader on a qLDPC challenge metric by approximately 48%. The agent also designed a simulated robotic hand mechanism that is 28.6% faster than a reference design.
On the open-source XiangShan RISC-V processor, the system found ten bugs and increased SPECint2006 per GHz by 5.79% while only increasing area by 0.3%. All results provided by Tencent are the company's own and have not undergone independent review.
