Hyra 2.0 Trains Models From Scratch, Beating ProteinMPNN and qLDPC Baselines

Key Info

Tencent introduced Hyra 2.0, the next evolution of its Hunyuan Research Agent, which moves beyond solving problems to building reusable tools, models, and workflows. The post highlights two models trained from scratch as evidence of that shift.

Highlights

• A 1.96M-parameter inverse folding network that runs 57x faster than ProteinMPNN.
• A qLDPC code achieving kd²/n about 48% above the qLDPC Challenge leader (non-CSS, weight ≤8).
• The agent's stated goal is producing reusable tools and workflows that accelerate future discoveries.
• Results are published alongside a perspective on the next chapter of AutoResearch.