Tencent says Hyra found a sharp construction for a 57-year sum-difference bound

Why it matters

A verified proof would show an AI research agent converting computational search into a general mathematical construction. It would also give Tencent a credible flagship application for Hyra beyond paper retrieval and routine analysis.

Tencent Hy said on July 30th that its Hyra research agent and Hy3 model found an explicit construction that makes a 1969 upper bound sharp in a long-running additive combinatorics problem.

The problem asks how much faster the sumset of a finite set of integers can grow than its difference set. For a finite, nonempty set $A$, the sumset $A+A$ contains every value $a+b$ for $a,b$ in $A$, while $A-A$ contains every difference $a-b$.

The target is the smallest universal exponent $c$ satisfying

$$ \frac{|A+A|}{|A|} \leq \left(\frac{|A-A|}{|A|}\right)^c $$

for every such $A$. Tencent says a theorem from 1969 established that $c=2$ is sufficient. Proving that 2 is the smallest possible value requires an infinite family of sets whose ratio approaches that upper bound.

Tencent's result graphic says Hyra and Hy3 produced an explicit family, denoted $A_K$, for which the corresponding exponent approaches 2 as $K$ tends to infinity. If the construction and proof withstand review, the result would close the gap between the known universal inequality and the best explicit lower-bound constructions.

A theorem claim, rather than another search score

The distinction between an explicit family and a high-scoring finite example is central to Tencent's claim.

Tencent's chart lists a series of previous results, including 1.0290 for a Marica construction, 1.0598 for a Freiman-Pigarev construction and 1.1259 for a Penman-Wells family. It also lists results from AI-assisted searches: 1.1440 for SimpleTES, 1.1449 after SimpleTES post-training and 1.2851 for a Codex system identified as using GPT-5.5 with human guidance.

SimpleTES researchers described their own result as a finite-set search driven by an exact evaluator. Their system maintained sum and difference multiplicity tables, then altered a structured arithmetic-progression backbone to improve the measured ratio. That produced a stronger individual construction, rather than an asymptotic family proving the universal exponent is optimal.

Tencent says Hyra crossed that second threshold. The stated result does not merely raise the numerical record. It supplies a family intended to drive the exponent all the way to 2.

The mathematical result should still be treated as Tencent's theorem claim. The X announcement presents the conclusion and a comparison chart. Independent verification requires mathematicians to examine the complete construction and proof.

Hyra gets a flagship case nine days after launch

Tencent announced Hyra-1.0 on July 21st, describing Hunyuan Research Agent as a system built for performance-driven research and engineering tasks. Tencent said Hyra can recursively improve its strategies through self-play, self-evaluation and user feedback.

The sum-difference result gives Tencent a concrete case study nine days after that introduction. It also moves Hyra into the same category of research agents as Google DeepMind's AlphaEvolve, which combines language models with automated evaluators to search for algorithms and mathematical constructions.

Both systems focus on problems where a proposed answer can be scored or checked automatically. That structure allows an agent to generate candidates, measure them, retain useful ideas and repeat the process at a scale that would be impractical for a single researcher. The harder step is converting a high-scoring computational object into a general construction with a proof. Tencent's announcement says Hyra made that transition here.

Hyra ran with Hy3, which Tencent officially released on July 6th after previewing the model in April. The open-weight model uses a mixture-of-experts architecture with 295 billion total parameters, 21 billion active parameters and a 256K-token context window. Tencent released the weights under the Apache 2.0 license.

Tencent has positioned Hy3 around reasoning, tool use and agent workflows, supported by a model-development loop that draws feedback from Tencent products. Hyra adds a research scaffold around those capabilities: the model proposes and analyzes candidate approaches, while evaluation and recursive search push the work toward a measurable objective.

The timing serves Tencent's larger model strategy. Open weights and low inference costs make Hy3 available to developers, while Hyra offers evidence that Tencent can build valuable systems above the base model. A verified solution to the sum-difference problem would give Tencent a stronger argument than another benchmark table: an AI-assisted mathematical result with a precise statement, a known upper bound and a proof that can be checked line by line.