Tencent Hy4 AI: Self-Improving Model Boosts Throughput 31.8%

Tencent has released Hy4, the first model in its Hunyuan lineup to participate in its own development. This 770-billion-parameter model was used to automate the optimization of its training methods, data strategies, and evaluation frameworks. Here's the thing: Tencent describes this as an early-stage recursive self-improvement loop. This process resulted in a 31.8% increase in end-to-end inference throughput. Human engineers oversaw this loop, incorporating results and making architectural decisions. What's interesting is that Hy4 optimized its own training pipeline and analyzed bottlenecks in its inference serving system. This included refining operator fusion and inter-GPU communication. The bottom line: Developers considering Hy4 need to understand this disclosure, what the model can and cannot do, and relevant legal requirements.