On November 2, Tencent's Hunyuan team released Hyra 2.0, a system designed to let AI conduct research on its own. The launch page lists seven results: two protein models, a molecular force field, a mathematical proving pipeline, a quantum error correction tool, a robotic hand, and a set of optimizations for Xiangshan, the open-source RISC-V processor.

Tencent's message is that the approach has changed. Hyra 1.0, released three months earlier, was still a matter of throwing repeated attempts at a single problem, and it did produce something: a week after launch, running on the open-source Hy3 model, it cracked a problem in additive combinatorics that had stood for more than fifty years, the optimal exponent for sumsets and difference sets. It was the first time a Chinese open-source model had solved an open mathematical problem fully autonomously.

Version 2.0 aims to go a step further: not just solve problems, but turn the methods used into tools that can be picked up the next time around. Tencent's analogy is that the better the tools, the stronger the system that builds them. That is the ambition. Seven results are the output. The question is how many of them anyone else can actually take and use.

Two Delivered in Full

The two most convincing results are both in proteins.

HyInv has just 1.96 million parameters and handles inverse folding: given a target structure, work backward to the sequences that will fold into it. Designing 128 sequences on a 402-residue backbone takes 50.6 milliseconds in total on a single H20, 57 times faster than ProteinMPNN, the industry standard. Design quality comes in at 0.875 by scTM, 0.013 higher.

The other, HyPLM-Flash, is a protein language model with 650 million parameters, trained on 726 billion tokens, 0.73 times what ESM-2 650M consumed. What does the smaller diet buy? A Spearman correlation of 0.409 on zero-shot ProteinGym tests, level with ESM-2 650M, ESM C 600M and E1-600M. Level with, not ahead of. On speed it is far quicker, roughly 130,000 residues per second, 1.9 to 2.7 times the three others.

Weights, code and technical reports for both are out, and they were benchmarked head-to-head against ten external baselines. But the report itself includes an important line: all results come from computational evaluation and case studies, with no wet-lab work. HyInv also handles only single-chain backbones, with multi-chain and membrane environments listed as next steps, and scTM is computed by a structure predictor, so it only approximates whether something will actually fold.

Five Fall Short

The two mathematics results are the most interesting. HyProver pushed the Petty equidistant point conjecture forward by one dimension, to five, where the previous human best was four. It also produced a counterexample to Yau's circle action conjecture, a seven-dimensional non-simply-connected manifold. That counterexample is the only item in the entire release carrying third-party authorship, with three mathematicians named. But it is worth being clear: the counterexample addresses the non-simply-connected case, and the simply-connected version remains open.

On quantum codes, HyQEC constructed a code with n=488, k=123, d=22, reaching an encoding efficiency of kd²/n of 122.0, roughly 48% above the previous leader on the qLDPC Challenge under non-CSS conditions with check weight no greater than 8. It also did some grunt work for the leaderboard: 178 of 187 entries received code distance certification, and in the process the upper bound on the former leader's code distance was tightened from 95 to 86, meaning that code was not as good as the leaderboard claimed. But Tencent says this toolset has not been open-sourced.

The remaining two are the thinnest. Xiangshan's SPEC CPU2006 geometric mean score per GHz rose 5.79%, with single-core frequency unchanged and area up only 0.3%, but that figure is estimated from a performance dashboard, equivalent to two or three months of official iteration, and not a controlled A/B test. The spiral-staircase hand started out about 11% slower than the best existing approach in DR5 simulation, and only pulled ahead by 28.6% after the control algorithm was fine-tuned within the same budget. There is no physical build. HyFF covers 83 elements and sits on the Pareto front against all 16 public models, yet the repository contains neither code nor a report.

Only the two protein models come with all three: weights, code and a report. The two mathematics results come with proofs and named co-authors, but the methods themselves have no reusable code. The quantum code tool is explicitly not open-sourced, and the chip and robotic hand have only self-evaluation and simulation behind them. Tencent framed this system around building tools, but the point of a tool is that someone else can pick it up and use it. What cannot be picked up is closer to a showcase.