GLM-5.2 tops the fine-tuning leaderboard, raising questions. The benchmark author clarifies: there was no distillation of Claude.

According to Dongcha Beating monitoring, the open-source model GLM-5.2 has topped the autonomous fine-tuning benchmark PostTrainBench, drawing criticism from scaling01 for lacking practical value. He noted that it is highly unusual for the model’s ranking to jump from 22nd place to the top within just a few months, and because the tests lack a hidden set, they can easily prompt agents to engage in leaderboard hacking for targeted optimization, making the resulting models difficult to deploy in the real world.

However, supporters counter that under the condition of being limited to 10 hours on a single H100 GPU, it is not realistic to require agents to finish a general fine-tuning run; targeted optimization is simply a common practice in machine learning. Public logs show that GLM-5.2 has a clear experimental logic: it can automatically collect data under different sampling assumptions, autonomously plan a complete pipeline that includes establishing performance baselines, fine-tuning, and using rejection sampling to filter data, and attempt to avoid overfitting in the chain-of-thought.

The greater value of this controversy is that the publicly available run trajectories—originally used to evaluate fine-tuning capability—unexpectedly pierced the industry rumor in China about “heavily distilling Claude.” After reviewing the GLM-5.2 logs, benchmark author Maksym Andriushchenko pointed out that the model has fundamental differences from Claude in data collection, strategy combinations, and decision pathways, with no signs of imitation or distillation. Public, transparent third-party benchmarks instead have become the most direct window for open-source large models to prove their original R&D strength.