Tencent Hunyuan releases Hyra 2.0, shifting AI from solving scientific research challenges to autonomously developing scientif

Tencent Hunyuan on November 2 launched Hyra 2.0, an AI research system that shifts from directly solving scientific problems toward building reusable scientific models, research workflows, and software tools in-house. The system uses a three-tier Agent architecture (Overall, Mission, Worker) supporting long-horizon task management, parallel experiments, independent verification, and accumulated research experience. Tencent disclosed that Hyra 2.0 autonomously trained a protein design model, HyInv, which generates sequences 57x faster than ProteinMPNN, and HyPLM-Flash, a 650-million-parameter protein language model that is the fastest in inference among 12 compared models. The system also reported gains in quantum error correction, robot structure design, and chip R&D, including a 5.79% improvement in the geometric mean SPEC CPU2006 score per GHz for the open-source RISC-V Xiangshan processor. Results will be progressively opened to the community for validation and reuse. Watch for adoption of the released models and tools by the research community and further disclosures on the chip and quantum workstreams.