Tencent's Hyra 2.0 advances toward AI autonomous research: training models from scratch, designing robots, and even participating in the development of next-generation AI.

Beating AI News: Tencent’s Hunyuan has released research agent Hyra 2.0, shifting its focus from trial-and-error for individual problems to developing reusable scientific models, algorithms, and tools. To enable AI to conduct research over days or even weeks, Hyra 2.0 divides work into three agent layers: Overall (resource coordination), Mission (research planning), and Worker (experiment execution). Tasks are persistently stored; while waiting for training results, other research can proceed. Each conclusion must be accompanied by evidence such as code and logs, and is reviewed by an independent agent when necessary. Valid experiences are also accumulated as Skills for subsequent tasks. Tencent disclosed that Hyra 2.0 has trained two protein models from scratch: HyInv, a 1.96 million-parameter model that designs protein sequences based on target backbones, generating sequences 57 times faster than the widely used ProteinMPNN. In a warm-up test on a single H20 GPU, generating 128 sequences of ~400 residues takes only ~50 milliseconds. The other model, HyPLM-Flash, has 650 million parameters and was trained on 726 billion tokens. The weights, inference code, and technical reports of both models are now public. Hyra 2.0 also participated in optimizing the open-source RISC-V processor Xiangshan, boosting its SPEC CPU2006 performance per GHz by 5.79% while only increasing chip area by 0.3%. It also designed a five-finger "spiral ladder hand" with incrementally longer fingers; initially, this hand was ~11% slower than the best existing design, but after training with the same budget control algorithm, its simulated rotation speed increased by 28.6%. Currently, Hyra 2.0 is also involved in pre-training, post-training, and inference system optimization for Tencent’s next-generation Hunyuan model, aiming to help develop a stronger next-generation AI by improving R&D tools.