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

Beating AI News Flash: Tencent Hunyuan releases research Agent Hyra 2.0, expanding its research focus from repeated trial and error on individual hard problems to developing scientific models, algorithms, and tools that can be reused over the long term.

To enable AI to continuously conduct research for days or even weeks, Hyra 2.0 divides work among three layers of Agents—Overall, Mission, and Worker—responsible for resource coordination, research planning, and experiment execution, respectively. Tasks are persistently saved, and other research can proceed while waiting for training results. Each conclusion must be supported by evidence such as code and logs, and when necessary, handed to an independent Agent for review. Effective experience is also distilled into Skills for use in subsequent tasks.

Tencent disclosed that Hyra 2.0 has already trained two protein models from scratch. Among them, the 1.96 million-parameter HyInv can design protein sequences based on target backbones, generating 57 times faster than the commonly used ProteinMPNN. In a warm-up test on a single H20 GPU, generating 128 sequences of about 400 residues each took only about 50 milliseconds. The other model, HyPLM-Flash, has 650 million parameters and used 726 billion training tokens. The weights, inference code, and technical reports of both models have all been made public.

Hyra 2.0 also participated in optimizing the open-source RISC-V processor "XiangShan," increasing SPEC CPU2006 performance per GHz by 5.79% while increasing chip area by only 0.3%.

It also designed a "spiral staircase hand" with five fingers of increasing length. This hand was initially about 11% slower than the previous best design, but after training with a control algorithm under the same budget, its simulated rotation speed became 28.6% faster instead.

Currently, Hyra 2.0 has also participated in the pretraining, post-training, and inference system optimization of the next-generation Hunyuan model, attempting to let AI help train a stronger next-generation AI by improving R&D tools.