Tencent lets AI build its own tools for protein research and mathematics

The Chinese tech company Tencent has presented a new version of Hyra, an AI system that does research on its own. Hyra does not just answer a question. It thinks up its own experiments on the computer, runs them, judges the result and then makes a better attempt. The first version appeared in July 2026 and focused mainly on better solutions for separate problems, such as mathematical puzzles and packing shapes as tightly as possible.

With Hyra 2.0, which Tencent announced on 2 November, the goal changes. Hyra now has to leave behind tools that can be reused later, mainly trained AI models for scientific work. A later project then does not have to start from zero. To handle research that takes days or weeks, Hyra 2.0 divides the work over three layers: the top one distributes the computing power, the middle one makes a plan for each assignment and the bottom one runs the experiments.

Public models for proteins and molecules

Part of these tools has been on GitHub since 2 November, the site where programmers share code with each other. It includes the models themselves, the code to use them and technical reports. Anyone may use them, including for commercial work.

The first model, HyInv, helps to design proteins. A protein is a chain of building blocks, called amino acids, that folds up by itself into a fixed shape. That shape determines what the protein does. Anyone who wants to make a new protein, for example for a medicine, often starts with the shape they want and then looks for a sequence of building blocks that folds up in exactly that way. HyInv suggests such sequences. Researchers often want to compare thousands of suggestions before they make a few in the lab, so speed matters.

According to Tencent's own report, HyInv makes 128 suggestions for a protein of about 400 building blocks in 0.05 seconds. ProteinMPNN, a widely used program for the same work, takes almost 3 seconds for this on the same chip: HyInv is 57 times as fast. Tencent checked the quality with programs that predict how a protein folds up. In those tests, HyInv's suggestions matched the desired shape slightly better on average than those of ProteinMPNN.

The second model, HyPLM-Flash, is a language model for proteins. A language model for text learns which words fit together in a sentence. In the same way, this model learns from 2.6 billion existing proteins which building block fits in which place. Researchers use such a model, for example, to predict whether changing one building block makes a protein work better or worse. According to Tencent, HyPLM-Flash performs as well as comparable models from other makers while being trained on less data: for some competitors only a fifth to an eighth as much. That saves computing time and therefore money.

Later on 2 November, five more models were added, together called HyFF. They predict how atoms attract and repel each other. With them you can imitate on a computer how molecules move, without working out every step with heavy physics calculations. According to Tencent, the fastest of the five is about as accurate as the comparable model UMA-S-1.2 from another maker, but 29 times as fast.

Two mathematical results

Tencent also published two mathematical papers. The first version of both proofs came from HyProver, another Tencent AI system that drafts mathematical proofs. Mathematicians at Tongji University in Shanghai then checked, improved and rewrote those proofs, and they are listed as authors of the papers. One of them, Kaiwen Yang, also works for Tencent. The next step is for mathematicians outside this group to assess the proofs.

The first paper is about a conjecture by the mathematician Petty. On a flat surface you can place three points that are all the same distance from each other: the corners of an equilateral triangle. Petty conjectured that in any space you can always place one point more in this way than the space has dimensions: three points in two dimensions, four in three, and so on. This would also hold if you measure distance in another way than with a ruler, for example by adding up the steps along the straight streets of a city grid. For three and four dimensions this had already been proved, the last time in 2005. The new paper proves it for five dimensions: there, six such points can always be found.

The second paper disproves a conjecture by Shing-Tung Yau, one of the best-known geometers in the world. Yau conjectured that every closed space that is curved like a sphere everywhere has a rotational symmetry: you can make it turn neatly, like a sphere around its axis. To disprove such a conjecture, one space for which it does not hold is enough. The mathematicians and HyProver found one, in seven dimensions. For a more strictly defined kind of space, the question is still open.

Hyra also helps with Tencent's next AI model

Tencent has its own large language model, Hy. The first version of Hyra ran on Hy3, the current model in that series. When Hyra started, Tencent already said that new Hy models and Hyra would grow together. According to Tencent, Hyra 2.0 is now helping with the next generation of Hy, both in training it and in making the model run faster. A better Hy model can then in turn produce a better Hyra. According to Tencent, the results of this work will follow on GitHub later.

Tencent also reports that Hyra 2.0 has come up with a new way to protect quantum computers against errors. Quantum computers calculate with qubits, and these make errors easily. So you spread the information over a larger group of qubits, which makes it possible to repair an error. Qubits are difficult and expensive to build, so the more of that group you can really use for calculating, the better. In Hyra's design, 123 of 488 qubits are left over for calculating, about one in four. According to Tencent, the design thereby scores about 48 per cent higher than the previous leader on a public ranking of the Unitary Foundation, an organisation that supports open software for quantum computers. Tencent has not yet put this design on GitHub.

The next step: checking the figures

The figures for speed and quality come from Tencent's own tests. Because the models for proteins and molecules are public, other researchers can now test them themselves. Tencent does not share the code of Hyra itself, nor exactly how the work went. As a result, nobody outside can see how many people helped, for example in choosing the problems and picking the best results, note the AI news sites AlphaSignal and remio. The public models do show what has been delivered. How independently Hyra did this, Tencent may still show later.

What does this mean for the road to the singularity?

Hyra 2.0 shows how AI can take part in scientific research and in doing so deliver tools that other researchers can start using straight away. The protein models can now be used by anyone, and two mathematical results have been checked and rewritten by mathematicians.

The most forward-looking part is the announcement that Hyra is helping with the model on which later versions of Hyra can run. Such a loop, in which AI helps to make better AI, was described at the end of September 2026 by a group of 22 researchers, among them the AI pioneers Geoffrey Hinton and Yoshua Bengio, in a research paper that appeared through an AI programme of the University of Cambridge. According to them, such a loop can shrink years of progress to months. They asked governments to keep track of how far companies are already automating their own AI research. At Tencent, this loop is still an announcement. The results that the company promises on GitHub will soon show how far along it is.

What is the technological singularity? Read the explanation →