Tencent Hunyuan released Hyra 2.0 on November 2, 2026 — a new generation of research agent (Hunyuan Research Agent) that does not operate on a "one task, one run" formula but instead accumulates reusable scientific models, workflows, and tools: the system stores acquired lessons in a shared knowledge base called "Books" and converts them into skills for future tasks. As part of this work, the agent independently trained specialized models for protein engineering and openly published codes, weights, and technical reports, allowing key release claims to be verified without access to Tencent's internal systems.

image
image

What happened

The Tencent Hunyuan team released the second version of its research agent, Hyra. The previous version, Hyra 1.0, was released on July 21 and operated on a minimal cycle of "propose solution — run experiment — evaluate metric." Hyra 2.0 introduced a three-tier hierarchy of Overall, Mission, and Worker agents, and the system is connected to Tencent's real research environment: the corporate messenger WeCom, a computing platform, and code repositories. During its operation, the agent independently trained a set of specialized models and published their weights for community verification. The list of artifacts includes the protein inverse folding model HyInv with 1.96 million parameters, which generates structures approximately 57 times faster than ProteinMPNN; the protein language model HyPLM-Flash with 650 million parameters featuring INT8 inference at a speed of approximately 130,000 residues per second on a single NVIDIA H20 card; and five HyFF force fields. Separately, the agent presented mathematical results: a counterexample to Yu's hypothesis on circular action on a 7-dimensional manifold and a proof of Petty's hypothesis in dimension 5. The publication covers both versions of the system: codes, weights, and technical reports for Hyra 1.0 and 2.0 are open in the Tencent-Hunyuan/Hyra-results repository, and the architecture is described on the project's official page.

Context

The background for this release is the standard operating scheme for research agents, known in the industry as AutoResearch: each run optimizes one task within a set compute budget, and the experience gained disappears with the completion of the run. Hyra 2.0 challenges this logic: lessons are stored in a shared "Books" database and then returned to work as skills, so the marginal cost of each subsequent task decreases rather than resets, and the agent becomes richer over time instead of starting from a blank slate each time. A second important shift is the execution environment: most agent publications live in an isolated sandbox, whereas Hyra 2.0 works in conjunction with the company's industrial environment, making it an attempt at a working pipeline rather than a showcase demonstration.

Why this matters for the industry

For the industry, Hyra 2.0 is valuable not so much for individual models but for a reproducible architectural template: multi-level agent orchestration, a cumulative memory layer where a lesson becomes a skill, and the publication of weights for independent verification. The economic logic changes: each run leaves behind reusable models, pipelines, and lessons, so compute investment becomes an asset rather than an expense, and Tencent's internal economics are being restructured right now. Teams building agent systems receive a concrete reference for how to organize a cycle with independent verification and knowledge accumulation, rather than just optimizing a single metric. If the open artifacts pass independent verification, a wave of imitation is likely: laboratories will begin publishing "accumulated" tools and lesson databases, open-source forks of this pattern will appear, and the methodology for benchmarking research agents will come under pressure — evaluation will need to assess not a single run, but the system's accumulated capital.

Why this matters for users

For readers, the release is convenient because almost everything can be verified without Tencent's permission: the Tencent-Hunyuan/Hyra-results repository is open and contains codes, weights, and technical reports for both versions. Users can download HyInv weights and compare it with ProteinMPNN on their own hardware, check HyPLM-Flash throughput based on the stated performance on a single NVIDIA H20, and connect HyInv, HyPLM-Flash, or HyFF to their own prototypes — from protein design to molecular modeling. For those building their own AI agent, the official Hyra 2.0 page provides a working architectural reference, and the "Books" database serves as a clear example of how to set up a cumulative knowledge layer in one's own product, where every developed lesson remains in the system.

What is currently unknown / limitations

All impressive figures in the publications are speed-based: ~57× compared to ProteinMPNN for HyInv and high throughput for HyPLM-Flash in INT8. Model accuracy metrics (e.g., sequence recovery for inverse folding) are not present in the sources, and generation speed alone does not prove quality. The mathematical claims — a counterexample to Yu's hypothesis and a proof of Petty's hypothesis in dimension 5 — are either correct or not, but no information about formal verification or external peer review is provided in the source materials. Finally, Hyra 2.0 remains an internal Tencent tool without a public API, pricing, or SLA, and the claimed "Books → skills" accumulation effect is currently a hypothesis: it needs to be confirmed outside Tencent's infrastructure and on external data.

Sources

Author

Look at AI, editorial team