On November 2, Tencent Hunyuan released Hyra 2.0, an agentic framework for scientific research, and the idea behind it has changed: instead of having AI solve one hard problem directly, agents are sent to develop reusable scientific models, workflows, and software tools. When research has to run for days or even weeks, a single question-and-answer exchange is no help; what matters is a system that can push experiments forward on its own and accumulate tools. Hyra 2.0 bets on the second kind of capability.
What each of the three agent layers handles
Hyra 2.0 splits the work into three layers: Overall handles resource coordination, Mission handles research planning, and Worker handles experiment execution. The three layers of agents connect to WeChat Work, computing platforms, and code repositories, working in a real research environment rather than running up scores on closed test problems. The advantage of the division of labor is clear responsibility: a wrong plan can be traced back, a failed experiment can be redone, and a finished tool can be kept for the next task.
Protein, quantum codes, and chips: what each result shows
Tencent Hunyuan published three kinds of results that have already been produced. In proteins, Hyra 2.0 trained two models from scratch, HyInv and HyPLM-Flash: HyInv has only 1.96 million parameters, can design protein sequences from a target backbone, and generates 57 times faster than the widely used ProteinMPNN; HyPLM-Flash has 650 million parameters and used 726 billion training tokens. The weights, inference code, and technical reports for both models have been made public. In quantum computing, the HyQEC tool constructed a quantum code whose encoding efficiency is about 48% higher than the previous best result in the qLDPC Challenge. In chips, Hyra 2.0 helped optimize the open-source RISC-V processor Xiangshan, raising SPEC CPU2006 performance per GHz by 5.79% while adding only 0.3% in chip area. It also designed a five-fingered "staircase hand" whose fingers increase in length.
What these results share is reusability: the models, tools, and design methods stay behind, so the next research project can build on them instead of starting from zero every time. Hyra 2.0 is also already taking part in pretraining, post-training, and inference-system optimization for the next generation of Hunyuan models, which amounts to AI improving research tools in order to help train stronger next-generation models.
One caution, though: the figures above all come from Tencent Hunyuan's release, and existing coverage does not say anything about independent reproduction or peer review. Hyra 2.0 is aimed at research teams, not a product ordinary users can download and use directly. What is worth watching next is whether the published models and code can be reproduced by outside teams, and whether the framework works just as well in other research settings once it leaves Tencent's internal environment.
