HyInv (Tencent Hyra)
HyInv (Tencent Hyra) is a protein structure & design tool whose code is published by Tencent-Hunyuan on GitHub, first public in 2026. Tiny 1.96M-parameter ProteinMPNN-style inverse-folding model (backbone to sequence) trained with structure-reward RL by Tencent's Hyra 2.0 agent; CPU/GPU, fixed-position design, fused Triton fast path.
| Category | Protein structure & design |
|---|---|
| Developer (GitHub) | Tencent-Hunyuan |
| Input | Protein backbone (N, CA, C, O), single or multiple chains |
| Task | Inverse folding (sequence design) |
| Interface | CLI and Python package |
| Website | github.com/Tencent-Hunyuan/Hyra-results/tree/main/AI4Science/hyinv |
| GitHub | Tencent-Hunyuan/Hyra-results |
| First public | 2026-11-02 (source) |
| Code license | MIT |
| Weights license | bundled weights ship in the MIT-licensed package |
| What this allows | Permissive (commercial OK) Use, modify and ship it inside closed-source commercial products; keep the license text and copyright notices. What do these mean? |
| License source | github.com/Tencent-Hunyuan/Hyra-results/blob/main/AI4Science/hyinv/README.md · checked 2026-11-02 |
| Pricing | Open source |
| Tags | #inverse-folding#design#reinforcement-learning |
| Last checked | · links verified |
Can I use HyInv (Tencent Hyra) commercially?
Yes. The recorded license is permissive, so closed-source commercial use is allowed if you keep the notices.
- Code: MIT
- Weights: bundled weights ship in the MIT-licensed package
- Source: github.com/Tencent-Hunyuan/Hyra-results/blob/main/AI4Science/hyinv/README.md · license checked 2026-11-02
What the license types mean · not legal advice.
Specs
| Task | Inverse folding (sequence design) |
|---|---|
| Modalities | Protein backbone (N, CA, C, O), single or multiple chains |
| MSA required | — |
| Open weights | Yes (in repo, 7.8 MB safetensors) |
| GPU need | Optional (runs on CPU or GPU) |
| GitHub stars | 177 |
| Interface | CLI and Python package |
Compare with…
Pick 1–3 other Protein structure & design tools to compare side by side with HyInv (Tencent Hyra).
Related
DeepMind's open-source protein structure prediction model (monomer and multimer).
Structure prediction of proteins, nucleic acids, ligands and their complexes; hosted web server plus open code, weights on request for non-commercial use.
Open biomolecular interaction models (Boltz-1, Boltz-2) for complex structure and binding-affinity prediction.