Hyra 2.0

Hyra 2.0 is an automated research system developed by the Hyra Team at Tencent Hunyuan. A research post dated November 1, 2026 describes it as a successor to Hunyuan Research Agent, or Hyra, focused on developing reusable scientific models, software tools and research workflows. Tencent reports using it within its own research environment, with human approval for consequential resource and coordination decisions.^[1]

Its public results include protein models, molecular-simulation software, a single-cell embedding model and quantum-code verification tools. These artifacts provide specific outputs that researchers can inspect. They do not amount to a public release of every part of the research agent or establish that it can reproduce arbitrary scientific discoveries autonomously.^[1][2]

Research objective

The team distinguishes solving an individual problem from improving a tool that can support many later problems. A research agent might, for example, optimize a simulation program, train a scientific model or create a verification workflow, then use that output in subsequent experiments. The proposed benefit is reuse of a completed research artifact across multiple tasks.^[1]

This extends the role of large language models in AI for Science beyond producing an answer or searching over candidate solutions. The team's stated objective includes the design, implementation, training, evaluation and deployment of new models. Its description of future cycles in which better research tools help develop better underlying models is a research direction, rather than evidence of unrestricted recursive self-improvement.^[1]

Agent organization

Hyra 2.0's agent harness divides work among three levels. These names describe roles in the system, rather than three separately released language models.^[1]

Different workers can use different models. A Mission can pursue exploration, implementation and review in parallel. Tasks enter a persistent queue, and a scheduler starts workers according to priorities, dependencies and resource limits. Waiting for a training run or an approval does not require every other task to stop. Agents can also set a later wake-up point to resume work after an experiment or review checkpoint.^[1]

The team's environment connects the system to WeCom for communication, a compute platform for experiments and code repositories for implementation. Large compute expenditures and operations that synchronize shared knowledge require the relevant human approval.^[1]

Evidence, memory and review

A Research Board records ideas, results, evidence and important decisions. Workers reporting completion must supply evidence such as a code revision, experiment outputs, logs or an execution trace. A Mission can inspect the underlying record instead of relying only on the worker's summary.^[1]

When independent internal checking is required, a new Review Worker receives the original problem and evidence. The design aims to reduce the influence of the implementing worker's conclusion on the reviewer. Such review remains a component of Tencent's research process; it is not equivalent to external replication or journal peer review.^[1]

The system separates project coordination from an individual agent's working memory and shared reusable knowledge. Its shared knowledge store, called Books, records methods, reasons, conditions of use and supporting evidence. The team distinguishes observations from independently reproduced findings and from conventions that require human confirmation. Project reviews can turn reusable methods into skills for later work.^[1]

Released scientific models

The companion Hyra-results repository contains selected outputs, with model definitions, inference packages, weights and technical reports. The following are documented releases within that repository.^[2]

The protein-model reports attribute their training to Hyra 2.0. HyInv uses a backbone-conditioned graph network and structure-reward reinforcement learning. Its optimized sampler generates candidates through drafting and rewriting. HyPLM-Flash is a masked language model trained on 726 billion tokens.^[3][4][5][6]

The reports and software documentation specify important conditions for their speed figures. HyInv's BF16 path generates 128 designs for a 402-residue backbone in a reported 50.6 milliseconds on one NVIDIA H20, excluding compilation and graph-capture overhead. That path is numerically different from reference FP32 sampling. HyPLM-Flash reports approximately 130,000 residues per second on one H20 with its INT8 engine. These are the developers' measurements under their specified configurations.^[3][4][5][6]

HyFF's comparisons use the OMol25 validation set for accuracy and a 2,586-atom solvated system on one H20 for molecular-dynamics timing. Its documented packages support NVE, NVT and NPT simulation. HyCell's released evaluation instead concerns frozen cell embeddings on seven atlases excluded from pretraining. Those evaluations measure different scientific tasks and should not be combined into a general score for Hyra itself.^[7][8]

Quantum-code tools and mathematical work

HyQEC searches for and verifies the distances of quantum stabilizer codes. Its release includes a prover and a separately written checker for branch-and-bound coverage certificates. The documentation distinguishes checked SAT proofs, independently replayed coverage certificates and search records that only the original prover can rerun.^[9]

Its public, budget-limited sweep reports exact distances for 165 of 184 non-CSS codes from an October 16, 2026 leaderboard snapshot. The technical report's 178-of-187 result uses a later snapshot and longer research runs. These are different experiments. The README also states that evidence for some newly constructed codes has not yet been released, so those claims cannot all be checked from the downloadable package.^[9][10]

The repository additionally publishes mathematical manuscripts, including a proposed dimension-five result for Petty's equilateral-set conjecture and a proposed counterexample to the unrestricted form of Yau's circle-action conjecture. Its index explicitly says the latter construction is not simply connected and leaves the simply connected version open. Publication in the team's artifact repository identifies the claimed result and its scope, rather than establishing acceptance by the wider mathematical community.^[2]

Robotics and processor-design experiments

The launch article describes a five-finger robot-hand design evaluated in an in-hand rotation simulation. After the same policy fine-tuning budget, Tencent reports 3.277 radians per second, 28.6% above its similarly fine-tuned reference design in the DR5 setting. The initial design had performed worse before fine-tuning. The article explicitly limits this result to simulation and says it has not undergone physical-robot validation.^[1]

Tencent also reports work on the XiangShan RISC-V processor, including performance tuning, fuzz testing and a cache-coherence counterexample found with TLA+. Its claim that the performance change represents two to three months of ordinary development is an estimate based on a historical project dashboard, not a controlled comparison of researcher productivity.^[1]

Availability and limitations

The Hyra-results repository is a collection of research outputs. Its contents do not, by themselves, provide a complete installer, the full orchestration configuration or a replay of the internal research environment described in the launch article. Availability must be checked for each artifact, rather than inferred from the existence of an open repository.^[1][2]

Licenses also differ by artifact. The repository uses Apache License 2.0, while the released HyInv and HyPLM-Flash packages state MIT terms. HyCell documents attribution obligations for its training and test data. Users need the applicable notices for the package, model and data they actually use.^[2][3][5][8]

The evidence supports specific developer-reported experiments and inspectable outputs. It does not establish general scientific reliability, measured gains across all subsequent research projects or the removal of human direction. Computational validation, physical experiments and external scrutiny remain distinct from the system's own review procedures.^[1][9][10]

Continue reading

• HunyuanHunyuan is Tencent's brand for its foundation models, an umbrella covering large language models, video generation, image synthesis, 3D asset creation, and multimodal reasoning.Mentioned twice above
• PProtein language modelA protein language model (pLM) is a neural network trained on raw amino acid sequences using the same self-supervised objectives that power text models, with individual amino acids serving as tokens instead of…Also in AI for Science
• Harness (AI)Harness, in artificial intelligence, is the software scaffolding that wraps a machine learning model, most often a large language modelAlso in AI Agents
• HHy4Hy4 is a large language model in Tencent's Hy (Hunyuan) family, released through hosted services on November 4, 2026.Also in Chinese AI
• DDream-RSIDream-RSI (Recursive Self-Improvement through Evolving Worlds) is a framework for improving the exploration strategy of an AI discovery agent, described in a preprint posted to arXiv on 14 September 2026 by a…Also in AI for Science
• AAire (AI lab)Aire is a stealth-mode artificial intelligence research company founded in 2026 by former researchers from Meta and OpenAI.Also in AI for Science

More to read

• Inherent (AI lab)
• Physical Superintelligence
• Academic Research
• Discovery Loop
• Biohub
• OpenAI Navier-Stokes Proposed Solution

Topics:AI AgentsAI ResearchAI for ScienceChinese AI

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14Hyra Team. Hyra 2.0: The next step for automated research. Tencent Hunyuan research article, dated November 1, 2026. Chinese original; accessed November 4, 2026.
2. ^1 ^2 ^3 ^4 ^5Tencent Hunyuan. Hyra-results repository. Research artifact index, release news and repository license. Accessed November 4, 2026.
3. ^1 ^2 ^3Hyra Team. HyInv release documentation. Model architecture, inference modes, requirements and package license. Accessed November 4, 2026.
4. ^1 ^2Hyra Team. HyInv Technical Report. Report dated October 14, 2026.
5. ^1 ^2 ^3Hyra Team. HyPLM-Flash release documentation. Inference engine, precision modes and package license. Accessed November 4, 2026.
6. ^1 ^2Hyra Team. HyPLM Technical Report. Report dated October 12, 2026.
7. ^Hyra Team. HyFF release documentation. Model sizes, evaluation conditions, molecular-dynamics interfaces and requirements. Accessed November 4, 2026.
8. ^1 ^2Hyra Team. HyCell release documentation. Architecture, evaluation, data and attribution. Accessed November 4, 2026.
9. ^1 ^2 ^3Gao, Minbo. HyQEC release documentation. Proof-evidence levels, snapshot differences and released resources. Accessed November 4, 2026.
10. ^1 ^2Gao, Minbo. HyQEC: Constructing new quantum LDPC codes and certifying their distances. Hyra Team, Tencent Hunyuan, technical report dated November 3, 2026.

Show all 10 references