# Tencent's Hyra Agent Reports Beating the Historical Best on 29 of 55 Open Mathematics Problems

The system runs a generate, evaluate and revise loop on top of Tencent's open Hy3 model, and the company has published its demonstration artifacts on GitHub.

- Published: 2026-08-04T06:03:13.954Z
- Canonical: https://polylog.news/ai/2026-08-04/tencent-s-hyra-agent-reports-beating-the-historical-best-on
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [Polylog editors](https://polylog.news), [Tencent Cloud](https://www.tencentcloud.com/techpedia/144774?lang=en)

Tencent's Hunyuan team released Hyra-1.0 on July 21, a research agent built on the open-weight Hy3 model, which carries 295 billion total parameters with 21 billion active. The design is deliberately plain. The agent generates a candidate solution, scores it against a verifiable objective, identifies what failed, and produces a revision. Tencent calls that loop recursive self-improvement, though the improvement occurs in the artifact the agent produces rather than in the model weights.

The reported results span three domains. Across 55 open mathematics problems, [Tencent says Hyra beat the previously best-known result on 29](https://www.tencentcloud.com/techpedia/144774?lang=en). It improved qubit routing efficiency by 44.4% over the standard SABRE algorithm, and designed an addition Transformer with 15 trainable parameters, 58.3% fewer than the public record. Russian-language AI coverage [reported that the agent helped researchers settle a combinatorics question that had stood for more than fifty years](https://t.me/ai_machinelearning_big_data/10637), a claim Tencent has not framed in those terms itself.

What makes this checkable is that Tencent open-sourced the demonstration artifacts in a public repository. Qubit routing and parameter-count records are the kind of claims that can be re-run rather than debated, which puts them in a different category from a benchmark score reported in a slide deck. Whether the mathematics results hold is a question for the specialists in each subfield, and no independent audit of all 29 has appeared.

The wider narrative around these systems has moved well beyond the evidence. One widely circulated post [argued that AI good enough to build better AI has already started a compounding loop and that the next eighteen months will be unlike anything before](https://t.me/aipost/7722). Hyra does not demonstrate that. It demonstrates that a fixed model, a verification signal and many attempts can search a solution space better than the published human record in narrow domains where the objective function is machine-checkable. That is a meaningful and repeatable capability, and it is not the same thing as a model improving itself.

## What this means

The pattern that works is a competent open-weight model wrapped in a loop with a hard verifier, which means the value is shifting from the base model to the search harness around it. That favors anyone with a cheap model and a good objective function, and it erodes the position of labs whose advantage rests on weight quality alone. Tencent gains distribution for Hy3 by demonstrating what the harness can do, and cloud providers gain because this style of work consumes enormous inference volume per result. The constraint is that it only works where correctness is machine-checkable, which covers algorithm design, formal proof and circuit optimization, and does not yet cover open-ended engineering judgment.

## What to watch

- Whether outside researchers reproduce the qubit routing improvement over SABRE from the published artifacts, since that is the easiest of the claims to test and the fastest signal on whether the rest holds.
- Whether Western labs publish comparable verifier-driven search results on open problems, which would show the technique generalizes rather than depending on Tencent's specific setup.
- Whether Hyra-style loops move to domains without a clean verifier, such as experimental biology or systems design. Success there would be the genuine capability jump, and failure would mark the boundary of the method.