Tencent's Hunyuan team released Hyra 2.0 on November 2. The notable part is not another benchmark record. It is a change of objective: instead of having AI solve problems, have it build the tools that solve problems.

From puzzle-solving to tool-making

Hyra 1.0's proof point was driving an open-source model to resolve a combinatorics problem in additive number theory that had stood for over fifty years. The mechanism was simple: give the model a hard problem with a clear pass/fail criterion, then let it iterate.

The ceiling of that approach is equally simple. Solve the problem, and the gain stops at the problem.

Hyra 2.0 proposes different economics. If the agent participates in building the scientific model, the research workflow or the software tool, the improvement can be reused by downstream tasks. Tencent calls it compounding intelligence.

How the three layers divide labor

Overall talks to the human researcher and aligns on goals.
Mission owns a single research question long-term, deciding which route to pursue.
Worker reads code independently, runs experiments, analyzes failures — and is allowed to challenge Mission's judgment.

That last clause is the interesting design choice. The most common failure mode in research is a wrong direction at the top with diligent execution underneath. Letting the worker push back installs an error-correction path.

Three separate memory stores

Hyra 2.0 keeps project state, layered agent memory and a shared knowledge base ("Books") in three distinct stores.

Separating them makes post-mortems legible: which facts were temporary to this project, which were accumulated by this agent, which are methods reusable by anyone. After a project review, transferable methods get packaged as skills that later tasks can call directly.

Two open-source models came out of it

Hunyuan also open-sourced HyCell, a single-cell representation foundation model, and HyPDE, a physics-field foundation model — both developed with Hyra 2.0 in the loop. They are themselves artifacts of the "AI builds tools" thesis.

How to judge whether this is real

AI for Science has a chronic problem: stunning demos, hard reproductions. Three tests separate substance from theater.

Can others use the output? Open weights, runnable code, reproducible protocols — that is hard evidence.

Does the gain transfer? A tool that only helps on the original task is puzzle-solving with extra steps.

Are failures documented? Research produces far more failure than success. A system that only shows wins should be discounted.

The direction is right. The bottleneck in science is usually not the answer to one question, it is the shortage of good instruments. But this path is slower and harder to validate, and it needs enough time to show that peers actually adopt what it produces.