Tencent's Hyra 2.0 research agent builds its own protein, force-field and proof tools and reports new maths results
Tencent Hunyuan says the agent settled Petty's equilateral-set conjecture in dimension five and built a protein design model 57 times faster than ProteinMPNN; the results have not been independently checked.
Why it matters
Hyra 1.0 aimed agents at single problems; version 2.0 has them build reusable models and tools, a shift toward automated research pipelines. Tencent lists a 1.96-million-parameter inverse-folding model, a family of five force fields and a quantum error-correction code it says beats the prior qLDPC challenge leader by about 48%. On Yau's circle-action conjecture it reports a counterexample to the unrestricted form, while the simply connected version stays open. Results and code are in an Apache 2.0 GitHub repository, so outside mathematicians and chemists can now test the claims.
Key facts
- Announced by Tencent Hunyuan on 2 November
- HyInv: 1.96 million parameters, 57 times faster than ProteinMPNN
- HyPLM-Flash: 650 million parameters, trained on 726 billion tokens
- Petty's equilateral-set conjecture settled in dimension five
- HyQEC code [[488,123,22]], about 48% better on a qLDPC metric
- Released under Apache 2.0 on GitHub