From the source
Tencent says Hyra 2.0 will involve general-purpose agents in developing and improving scientific tools and models, so one development effort can be reused in later research.
- The company frames the shift as a move from agents that solve individual problems to agents that build the tools other research depends on.
- Hyra 1.0, released three months earlier and powered by the open-source model Hy3, solved an additive combinatorics problem that had stayed open for more than fifty years, which Tencent calls the first fully autonomous solution to an open mathematical problem by a Chinese open-source model.
- Tencent argues that sampling until a solution appears does not compound, while improving a tool benefits every downstream task that uses it.
- It sorts tools into symbolic ones, such as search algorithms and verification programs, and parameterized ones, such as language models and scientific models for protein design.
- Tencent says Hyra worked with Hunyuan Hy4 preview to accelerate force-field models for molecular dynamics, and that building a new model end to end is harder than accelerating an existing one because feedback arrives slowly.
- The company hopes Hyra 2.0 can improve training infrastructure, data synthesis pipelines and inference systems, and train better Hy models that Hyra 2.0 can then use.
