Tencent Hunyuan has released Hyra 2.0, an autonomous research agent that creates and refines scientific models, workflows and software for use in later projects. Compared with its predecessor, it adds coordination for long-running tasks and retains outputs that subsequent research tasks can use.
Three layers for managing research
Hyra 2.0 divides work among three layers of agents. The Overall agent sets research goals and allocates resources, while Mission agents manage individual tasks and track their progress. Worker agents write code, run experiments and evaluate proposals in parallel isolated environments.
Parallel Worker agents can test different approaches at the same time, while Mission agents retain state between experiments. Tencent says this structure supports independent verification and research programs lasting several days.
Experiments build on stored results
Tencent describes the agent’s work as a repeated research cycle. A Context Agent first organizes previous results and supporting material; Proposal Agents then generate candidate solutions from a shared queue. After experiments run in isolated environments, evaluators provide feedback for the next iteration.
The Experience Bank stores code, logs, outputs and feedback for future runs. An asynchronous design allows the queue to be continually replenished while several agents conduct experiments.
The previous Hyra 1.0 used an inner loop to optimize solutions and an outer loop to revise evaluators. Tencent applied that design in more than ten scenarios, including AI research, scientific discovery and game design.
Results in proteins, quantum codes and mathematics
According to Tencent, the agent trained HyInv, a protein inverse-folding model with 1.96 million parameters. The company says its generation speed is 57 times that of ProteinMPNN. It also says Hyra 2.0 independently trained HyPLM-Flash, a protein language model.
In quantum error correction, Tencent says HyQEC constructed a [488,123,22] quantum code. The numbers represent block length, logical qubits and code distance. The company says its encoding-efficiency score was about 48% above the previous leading result in the qLDPC Challenge.
HyProver generated proofs for mathematical problems including Shing-Tung Yau’s circle action conjecture; Tencent says mathematicians verified the work. HyInv, HyQEC and HyProver are designed for repeated use on new inputs.
Another reported result is an optimization of the XiangShan RISC-V processor: Tencent says Hyra 2.0 increased its performance score per gigahertz by 5.79%. The company also says the agent is contributing to development of the next-generation Hy model.
Output releases and code availability
Hyra 2.0 outputs are planned for release on GitHub, but no date has been set. The public Hyra 1.0 repository already contains outputs and reproduction scripts under the Apache 2.0 license.
The agent’s source code, model weights and complete methodology are unavailable. Complete prompts, orchestration and records of human intervention have not been published either.
