Tencent has released Hyra 2.0, a scientific research agent that autonomously manages multi-week experiments across protein modeling, robotic hardware design, and chip optimization, marking a step beyond single-task AI agents toward systems that build reusable scientific tools1,2.
Hyra 2.0 uses a three-layer agent architecture — Overall, Mission, and Worker — responsible for resource coordination, research planning, and experiment execution, respectively. Tasks are persistently saved, allowing the system to continue other research while waiting for long-running training jobs.
The agent trained two protein models: HyInv, a 1.96-million-parameter inverse-folding model that designs protein sequences based on target backbones, and HyPLM-Flash, a 650-million-parameter protein language model trained on 726 billion tokens. HyInv generates results 57 times faster than ProteinMPNN. In a warm-up test on a single H20 GPU, generating 128 sequences of approximately 400 residues each took about 50 milliseconds. Tencent made the weights, inference code, and technical reports for both HyInv and HyPLM-Flash publicly available.
Hyra 2.0 designed the Xuantie Hand, a robotic hand with five fingers of increasing length arranged in a spiral pattern. After training with the same budget of control algorithms, the spiral hand's simulated rotational speed increased by 28.6%, though the design was initially about 11% slower than the previous optimal configuration.
On the open-source RISC-V processor Xiangshan, Hyra 2.0 improved SPEC CPU2006 performance per GHz by 5.79% while increasing chip area by only 0.3%.
The agent now supports pre-training, post-training, and inference system optimization for the next-generation Hunyuan model. That work follows Tencent's October 29 announcement of Hunyuan Hy4, which added visual understanding to the flagship model for the first time[1].
ANALYSIS The breadth of domains Hyra 2.0 operates across — from protein science to processor microarchitecture to robotic morphology — distinguishes it from narrower AI research agents that target a single discipline. Open-sourcing the protein model artifacts while folding the agent into Hunyuan's own training pipeline ties Tencent's scientific automation effort directly to its commercial model roadmap.