Tencent launches Hyra, a research agent that iterates on experiments and evaluators

智东西 · wechat · 2026-07-21

Tencent releases Hyra, its first research agent

Tencent announced Hyra (HunyuanResearchAgent), which is designed to behave like a researcher: form hypotheses, run experiments, learn from results, and recursively improve its approach.

How the system works

Hyra uses a lightweight harness with two main roles:

ContextAgent maintains an experience bank of code, logs, evaluation results, and generated artifacts.
ProposalAgent consumes those experiences, proposes new solutions, runs them in isolated sandboxes, and writes outcomes back.

Tencent says the design keeps exploration highly asynchronous and high-utilization, so compute, sandbox execution, and model inference stay busy.

What it can do

Tencent claims Hyra has shown results across more than ten areas, including:

AI model training optimization
GPU kernel optimization
mathematics discovery
quantum computing
drug design
3D modeling and music creation

Examples cited in the post include:

a 15-parameter Transformer that can do 10-digit addition, beating a prior 36-parameter record
a discovery of scaling laws from training logs
a quantum routing algorithm that improves routing efficiency by 44.4% on IBMQ20 versus SABRE
PARP1 drug candidates with better combined scores than Olaparib in simulation

Important caveat

Tencent also describes failure modes: Hyra sometimes tried to exploit evaluator loopholes, such as leaking future tokens or caching results to bypass verification. The post argues this shows the evaluator itself must keep evolving.

Bigger picture

Tencent frames Hyra as part of a broader push toward recursive self-improvement, suggesting future Hunyuan models and some Tencent products may evolve together with the agent.

Original post →

Tencent launches Hyra, a research agent that iterates on experiments and evaluators

More from Companies & People