On July 21, the Tencent Hunyuan team officially released Hyra-1.0 (Hunyuan Research Agent), an intelligent agent capable of recursive self-improvement and specifically designed for performance-oriented research and engineering tasks. The system follows the "The Bitter Lesson" design philosophy—keeping the framework as lightweight as possible and expanding the agent's action space to a large extent, using a simple and general loop framework to convert intelligence and computing power into real returns.

The core architecture of Hyra consists of three parts: the Context Agent maintains an experience library and continuously generates "inspiration" context to fill the task queue; multiple Proposal Agents take context from the queue, generate and run solutions in independent sandboxes; the entire system operates as an asynchronous producer-consumer pipeline, with resource utilization efficiently expanding over time. For tasks without an evaluator, Hyra can be upgraded to a two-layer loop, allowing solutions and evaluation mechanisms to evolve together, effectively preventing reward hacking.

QQ20260721-135303.jpg

In the AI for AI field, Hyra outperformed the best results reported in Recursive in three benchmark tests: it reduced the validation set BPB to 0.9015 in the NanoChat Autoresearch task; it shortened the time to achieve a validation loss of 3.28 on the deeply optimized NanoGPT Speedrun to 76.4 seconds; and after jointly optimizing 235 GPU kernels on SOL-ExecBench, the Mean SOL reached 0.771.

In the AI for Science field, Hyra selected 55 mathematical open problems from databases such as EinsteinArena, of which 29 set new historical records. The team also provided Hyra with monthly sunspot data from 1749 to 1932, and the recurrence formula discovered by Hyra achieved an R² of 0.77 on out-of-sample data spanning nearly a century. More breakthroughs include Hyra designing a Transformer that can perform 10-digit addition with only 15 trainable parameters, reducing the previous public record by 58.3%; its qubit routing algorithm improved routing efficiency by 44.4% on IBM Q20 compared to the classical SABRE algorithm; and in drug design, the candidate molecules generated by Hyra for the anti-cancer target PARP1 showed significantly higher drug-likeness scores than the marketed drug olaparib.

In the AI for Fun field, Hyra's Go-playing program ranked third among 730 participating programs on the Botzone platform; it can generate renderable 3D models based on a single 2D reference image; and it can also compose full harmonies for various instruments based on a melody, generating richly colored complete arrangements after seven rounds of evolution.

The Tencent Hunyuan team stated that these are still preliminary achievements. In the future, they will integrate Hyra into product systems, real AI R&D pipelines, and industrial scenarios, driving the evolutionary cycle of "scaffolding-data-model," enabling the next generation of Hunyuan models and Hyra to continue co-evolving.