Hyra 2.0

Hyra 2.0 is an AI research-agent system (AutoResearch) released by Tencent Hunyuan (Tencent Hy) on November 2, 2026. Where the first Hyra mainly pointed general-purpose agents at individual hard research problems, Hyra 2.0 has agents build and upgrade reusable scientific models, research workflows and software tools, so that the output of one project keeps paying off across many later tasks.

Tencent frames this as the "second half" of AutoResearch: moving from lottery-style sampling and repeated problem-solving to tool upgrades. Hyra 2.0 has already published early results in protein design, molecular simulation, quantum error correction, mathematical proof, robot structure design and processor engineering, and it is now helping develop Tencent's next-generation Hy models.

Architecture: a three-tier agent hierarchy

Tier	Responsibility
Overall	Talks to researchers; coordinates priorities, resources and key decisions
Mission	Owns one research question long-term; plans routes, schedules experiments, judges conclusions
Worker	Writes code, runs experiments and analyzes results in parallel isolated sandboxes

• Tasks sit in a persistent queue, so work continues while waiting on experiments or human approval — enabling research programs that run for days to weeks.
• A Research Board tracks progress and requires Workers to submit evidence (code versions, results, run logs); independent Review Workers re-check when needed.
• Project state, working memory and cross-task experience are stored separately; validated methods become reusable skills.

Published results (vendor-reported)

• HyInv — a 1.96M-parameter protein inverse-folding model trained from scratch; better structure agreement than ProteinMPNN and 57x faster generation.
• HyPLM-Flash — a 650M-parameter protein language model trained on 726B tokens; fastest inference among 12 compared models, 3rd average rank over 35 tasks.
• HyFF — an atomistic foundation model for molecular simulation in 5 sizes (2.3M–1.2B params), 83 elements, on the accuracy/speed Pareto frontier.
• HyQEC — built a [[488,123,22]] quantum code, ~48% above the previous qLDPC Challenge leader on the efficiency metric.
• HyProver — a proof workflow that completed proofs including Yau's circle-action conjecture, checked by mathematicians.
• XiangShan RISC-V — +5.79% SPEC CPU2006 score per GHz; fuzzing and formal verification found 10 issues, with fixes contributed to the open-source OpenLLC cache.

How open is it?

Artifacts (models, code, reproduction scripts) are being published on GitHub under Apache 2.0, but the agent implementation, weights and full orchestration methodology remain closed.

How to follow or use it

1. Read the official write-up — "Hyra 2.0: The Next Step for Auto Research" on Tencent Hy's research page (hy.tencent.ai/research) and the Tencent Hunyuan WeChat post.
2. Reuse the artifacts — watch Hyra's GitHub results repo for HyInv, HyPLM-Flash, HyFF models and HyQEC / HyProver outputs with reproduction scripts (Apache 2.0).
3. Borrow the architecture — the Overall / Mission / Worker hierarchy, persistent queues, evidence-based Research Board and Review Workers are a useful blueprint for long-horizon research or engineering agents.
4. Compare with peers — contrast it with DeepSeek's dsh-superagent (Router/Worker) and other AutoResearch / AI-Scientist systems to understand different approaches to long-task orchestration.

Note: the agent system itself is not publicly available; all figures are Tencent-reported and await independent verification on the released artifacts.