Tencent Hunyuan Hyra 2.0 lets research agents build up tools; commenters first ask about token costs

Tencent Hunyuan has released Hyra 2.0, calling it the next generation of the Hunyuan Research Agent. The official tweet puts the change plainly: from solving problems to building reusable tools, models, and workflows that accelerate subsequent discovery. The official post also provides a complete explanation of the next chapter of AutoResearch: https://hy.tencent.com/research/hyra2

The common demo for research agents used to be: give the agent a problem, it runs, and returns an answer or proof. Figure 2 places the two paths side by side: top left, researchers design and train a specialized AlphaFold model to predict protein structures; top right, a general research agent solves a math problem and obtains an answer or proof; in the middle is Hyra's path, where researchers provide goals and feedback, and an agent driven by a general model develops and upgrades reusable tools. The official announcement lists three categories: scientific models Hylv, HyPLM, and HyFF; the research workflow HyProver; and the software toolkit HyQEC. At the bottom of the figure is a sentence: one tool-development effort can bring benefits to many subsequent research tasks.

This positioning is worth noting. The output of problem-solving agents usually stays within a single task; change the problem and you start over. Hyra 2.0 wants to leave assets behind so that the next task does less repetitive work. Someone in the comments asks very specifically: how exactly is a reusable tool defined? Others worry that agents being able to check each other's sessions alone will push token costs up threefold. Whether reuse can cover development and maintenance costs is a question the launch event did not answer.

According to Pandaily, citing Tencent's release, Hyra 2.0 tuned the open-source Xiangshan RISC-V core, increasing the SPEC CPU2006 score per GHz by 5.79% with a 0.3% area increase; its fuzzing found 7 bugs that were not in the official issue tracker. This case is more verifiable than 'accelerating scientific research': the tuning and test results produced by the agent can land on concrete metrics.

Commenters are also urging Hunyuan 4, and some ask why the video model can no longer be accessed. These are not the same thing as Hyra 2.0, but they show that readers are sensitive to release cadence and product availability. For research agents to prove the value of reuse, the ultimate test is whether subsequent tasks really involve writing less code, running fewer experiments, and tuning fewer parameters.