Tencent Hunyuan releases Hyra 2.0, a research agent that builds reusable tools, under Apache 2.0

Tencent Hunyuan released Hyra 2.0 on November 2, the second version of its autonomous research agent, with a design aimed at building reusable research tools. Results and code are published in an Apache 2.0 GitHub repository, per a report from Pandaily.

Hyra 1.0 focused on solving individual problems, including settling a sum-versus-difference exponent problem with the Hy3 model. Hyra 2.0 shifts to producing tools that can be used again.

Architecture

Hyra 2.0 is a three-layer system of overall, mission and worker agents. They are coordinated through a shared research board and a persistent task queue. A dedicated review worker checks outputs, agents keep layered memories, and a shared knowledge base distills lessons into reusable skills.

Tools and results Tencent reports

Proteins

• HyInv: an inverse protein-folding model with 1.96 million parameters. Tencent says it runs 57 times faster than ProteinMPNN with better structural agreement, and designs 128 sequences in 50 milliseconds.
• HyPLM-Flash: a 650-million-parameter protein language model trained on 726 billion tokens. It was the fastest of 12 models compared and ranked third on average across 35 tasks, processing about 130,000 residues per second on one H20.

Force fields

• HyFF: a family of five force fields, from 2.3 million to 1.2 billion parameters, covering 83 elements. It sits on the accuracy-speed Pareto front against 16 models.
• Tencent says its Largo model posts the lowest energy error, 0.364 meV per atom. The suite simulated a 60,000-atom aquaporin system for 100 nanoseconds.

Mathematics

• HyProver settles Petty's equilateral-set conjecture in dimension five, raising the highest settled dimension from four.
• On Yau's circle-action conjecture, the GitHub write-up describes a counterexample to the unrestricted form: a closed seven-manifold that is not simply connected. The simply connected version remains open.

Quantum error correction

• HyQEC found a [[488,123,22]] code. Tencent says it beats the previous leader on a qLDPC challenge metric by about 48%.

Hardware

• Hyra designed a simulated robotic hand mechanism 28.6% faster than a reference design.
• On the open-source XiangShan RISC-V processor, it raised SPECint2006 per GHz by 5.79% with 0.3% more area, and found ten bugs.

Tencent also says Hyra is now helping train its next Hy models.

Verification status

All results are Tencent's own and have not been independently reviewed. Pricing and API availability are not part of the announcement; the release is code and results on GitHub under Apache 2.0.

What this means

The change from Hyra 1.0 to 2.0 is in the output: the system is meant to produce tools such as models, force fields and codes, not only answers to single problems. The reported range spans protein design, atomistic simulation, pure mathematics, quantum codes and processor design, so each claim needs checking in its own field.

Because the results are self-reported and unreviewed, the speed, accuracy and conjecture claims remain Tencent's assertions for now. The Apache 2.0 repository lets outside researchers inspect the code and the Yau counterexample write-up. The note that Hyra is helping train the next Hy models indicates Tencent is using the agent in its own model development, though the announcement gives no detail on how.