Tencent Hunyuan releases Hyra 2.0, shifting automated research from solving to building tools
Tencent Hunyuan's Hyra 2.0 splits research across Overall, Mission and Worker layers, producing the HyInv protein model at 57 times the speed of a common industry method and a record quantum code efficiency.
Tencent Hunyuan formally released the Hyra 2.0 AI research system on November 2, shifting its R&D focus from using AI to solve scientific problems directly toward letting general agents participate in developing and upgrading scientific models, research workflows and software tools.
Hyra 1.0 had previously driven an open model to solve an open problem in additive combinatorics left unresolved for more than fifty years, validating the approach of giving AI a problem with clear evaluation criteria and letting it iterate. Hyra 2.0 introduces a new harness architecture splitting research work across Overall, Mission and Worker layers: Overall interfaces with human researchers and manages project priorities, Mission owns a research question long term and is accountable for conclusions, and Worker independently reads code, runs experiments and can challenge Mission.
On results, HyInv is a graph neural network of just 1.96 million parameters addressing inverse folding in protein design, generating sequences 57 times faster than ProteinMPNN commonly used in the pharmaceutical industry. HyPLM-Flash is a 650 million parameter protein language model trained on a cumulative 726 billion tokens, the fastest inference among the 12 protein language models compared. HyFF is an atomic-level foundation model for molecular simulation offered in five sizes from 2.3 million to 1.2 billion parameters covering 83 elements, all on the accuracy-speed Pareto frontier.
HyQEC targets construction and verification of quantum low density parity check codes, designing the Split Pauli-CPM algorithm and rigorously verifying the quantum code [488,123,22], with an encoding efficiency metric of kd squared over n reaching 122.0, about 48 percent above the previous qLDPC Challenge leader, and search throughput 8 to 55 times that of the existing 8-thread CPU tool. The system also contributed to performance optimization and correctness verification for the Xiangshan RISC-V processor. Tencent said results will be published in a GitHub repository for community inspection.