Tencent’s Hunyuan team released Hyra 2.0, a research agent designed to run projects lasting days or weeks and develop reusable scientific models, algorithms and tools, according to Beating AI and BlockBeats.
Persistent research workflow
Hyra assigns work to Overall, Mission and Worker agents for resource coordination, research planning and experiment execution. Tasks remain saved while training runs, and conclusions are expected to include code or log evidence, with independent review when needed.
Models and engineering applications
Tencent said Hyra trained the 1.96 million-parameter HyInv and 650 million-parameter HyPLM-Flash protein models from scratch; their weights, inference code and technical reports were made public. The report also described contributions to the open-source XiangShan RISC-V processor and a robotic-hand design, while Hyra is participating in pretraining, post-training and inference optimization for the next Hunyuan model.
The performance figures and release details are company disclosures relayed by BlockBeats; this brief does not independently reproduce the benchmarks. Source: BlockBeats English report, published Nov. 2, 2026. Chinese source.