While Silicon Valley debates whether AI will ever master counting letters in a word, Tencent has decided to stop playing with chatbots. A preprint by Haowei Lin and Shanda Li recently hit arXiv, effectively closing the book on an additive combinatorics problem that has eluded both mathematicians and elite neural networks. The issue concerns set expansion constants—a fundamental challenge where Google DeepMind’s proprietary AlphaEvolve and the latest models from OpenAI and Anthropic all stumbled. The solution was delivered by Hyra, a Tencent research agent built on the open-source Hy3 model. This isn't just a "neat result"; it’s a public drubbing of general-purpose LLMs by task-specific autonomous systems.
The situation is almost comical. For over a year, industry leaders stormed a barrier in a problem requiring a link between the sizes of sums and differences of sets. The human record had stood since 2013 at 1.1259. Terence Tao brought the AlphaEvolve system to DeepMind, which in late 2024 produced... 1.1219. In other words, a result worse than what humans achieved twelve years ago. An avalanche of agents followed: Baidu’s LoongFlow, SimpleTES, and EvoMaster. They spent months squeezing out the fourth decimal place, reaching 1.1449. Then Hyra arrived and proved the correct answer is exactly 2. To understand the scale of the miss: every previous system tried to enumerate sets head-on, but the objects grow so rapidly they cannot be held in memory.
The Lean version of the proof includes an explicit example whose index differs from 2 by less than 10⁻⁹⁹⁹. Everyone was looking for a better number—when they should have realized the limit coincides with the ceiling of the classical inequality.
The result is verified in the Lean 4 formal language, which automatically shuts down any talk of hallucinations. For R&D departments, this is a critical signal: the era of waiting for chatbots to have a "eureka moment" is ending. The age of recursive self-improvement systems has arrived.
Anatomy of the Chinese bypass
The secret to Tencent’s success lies in radical pragmatism. Lin and his team realized that forcing AI to generate endless blocks of Python code was a dead end. They allowed the Hyra agent to propose constructions and proof sketches in plain text. The system ran in this mode for about 24 hours. The role of the "judge" selecting the best ideas was played by GPT-4o (or its iterations), but it served only as a navigator, not the source of truth. The intellectual synthesis of the new mathematical construction was performed by Hy3—an open-source Mixture of Experts (MoE) model with 295 billion parameters. While Western giants burn budgets in closed labs, Chinese engineers are building architectures where only 21 billion parameters are active from the total mass. This allows for heavy research processes with manageable CapEx.
Hyra (Hunyuan Research Agent) is not a chat interface; it is an Agent OS. One module generates hypotheses while others test them in isolated sandboxes. According to Tencent, the system broke historical records in 29 out of 55 open math problems. The fact that the Hy3 model is published under the Apache 2.0 license turns this success into a geopolitical shift. A tool of this caliber is now available to any player with the hardware, bypassing export bans on cloud access to closed APIs.
The end of the generalist era
For businesses, this case is a funeral march for the "one model for everything" concept. In zero-tolerance environments, generalists always lose to hierarchical agents. Renowned mathematician Thomas Bloom, who verified the work, acknowledged the result as correct and listed the AI as a full co-author: "Lin, Li, and Hyra." This is a rare instance where an algorithm stands alongside Carnegie Mellon scientists not as an assistant, but as the author of the core idea.
The cost of intellectual discovery in fundamental fields has begun to drop—not through infinite server scaling, but through proper agentic wrapping of open-source software. An approach combining specialization and formal verification has proven more effective than billion-dollar investments in "Closed AI" and its perpetual struggle against hallucinations. In the race for leadership, victory goes not to the heaviest models, but to the most rigorous idea-filtering systems.
Strategic context
The success of the Hunyuan Research Agent marks a shift from conversational AI to autonomous R&D systems. By treating AI as an "Agent OS" with specialized modules for hypothesis generation and verification, Tencent has demonstrated that open-weights models can outperform the world’s most expensive closed systems in high-reasoning tasks.