Tencent Hunyuan introduces Hyra 2.0, the next evolution of its Hunyuan Research Agent, which the company says moves from solving problems to building reusable tools, models and workflows. PANews, citing Tencent Hunyuan's official WeChat account, dates the launch to November 2, 2026, and describes Hyra 2.0 as an automated scientific research framework in which agents develop and upgrade reusable scientific models and tools.

Image shared by @TencentHunyuan on X
Image shared by @TencentHunyuan on X

Image credit: @TencentHunyuan on X

What changes from Hyra 1.0

Baidu Baike says Hyra-1.0 launched on July 21 as an agent capable of recursive self-improvement for performance-oriented research and engineering tasks. TechFlow, citing Tencent's announcements, says Hyra 2.0 uses a three-tier architecture of Overall, Mission and Worker agents. That structure supports long-term task management, parallel experiments, independent verification and accumulation of research experience. AlphaSignal reports that the update expands the agent's ability to manage long projects and preserve useful work.

Reported results

PANews lists these results from Hyra 2.0:

  • HyInv: a protein inverse folding model trained from scratch, with 1.96 million parameters, generating about 57 times faster than ProteinMPNN.

  • HyQEC: a quantum error correction tool that built a quantum code whose encoding efficiency metric is about 48% above the former top entry of the qLDPC Challenge.

  • HyProver: a mathematical proof workflow that completed proofs of problems including Shing-Tung Yau's circle action conjecture, which PANews says mathematicians verified.

  • XiangShan: a RISC-V processor whose score per GHz rose 5.79%.

TechFlow adds that Hyra 2.0 independently trained HyInv and the protein language model HyPLM-Flash, which has 650 million parameters and was the fastest at inference among 12 compared models. PANews reports that Hyra 2.0 is also helping develop the next-generation Hy model, and that related results will be published on GitHub, without a date.

Earlier results and where to check them

The Hyra-results repository on GitHub holds artifacts from Hyra 1.0 and lists its results against prior bests. One example is sums_diffs at 1.21079, against a prior best of 1.14489. The repository says the agent itself is not included and links to Tencent's launch post for the earlier version.

Our read: the clearest way to test these claims is the public artifacts. The Hyra 2.0 results are reported through Tencent's announcements and PANews, TechFlow and AlphaSignal, which repeat them.

Sources