Tencent Hunyuan Releases Hyra 2.0, Where Agents Develop Reusable Scientific Models and Tools
Reported by panews.io and Tech Times2 sources ↓
Tencent Hunyuan has released Hyra 2.0, a framework designed to have AI agents build scientific models and tools for repeated use. Tencent says its outputs include a protein model, a quantum error-correction code and mathematical proofs.
Key points
- Tencent says HyInv has 1.96 million parameters and generates sequences 57 times faster than ProteinMPNN
- Tencent says HyQEC's encoding-efficiency metric is about 48% above the previous qLDPC Challenge leader
- HyProver completed proofs, including one involving Shing-Tung Yau's circle action conjecture, which mathematicians verified
- Hyra 2.0 raised the XiangShan RISC-V processor's score per GHz by 5.79%, according to Tencent
- Tech Times reports Hyra 2.0 is participating in training Tencent's next-generation Hunyuan model
Sources
Details
Tencent describes Hyra 2.0 as a shift from solving individual scientific problems to producing models and tools that can be reused. PANews reports that the framework trained HyInv from scratch and built HyQEC, a quantum error-correction tool; it also reports that Hyra improved the XiangShan processor's score per GHz by 5.79%.
The performance figures are Tencent's claims, Tech Times reports, and it says independent replication has not been published. Tech Times says HyInv and another protein model, HyPLM-Flash, are available on GitHub with weights, inference code and technical reports; it also reports that HyQEC was entered in the Unitary Foundation qLDPC Challenge leaderboard.
PANews reports that HyProver completed proofs including Shing-Tung Yau's circle action conjecture, and that mathematicians verified the proofs. Their names have not been released, according to Tech Times. Tencent says related results from Hyra's work on the next-generation Hunyuan model will be made public on GitHub.
Written from coverage as of
Images
- panews.io
- Tech Times
Stocks in this story
Change since open: 20 Nov to 21 Nov 2026
Last refresh