GLM5.2 this wave is pretty tough, didn't crash in 84 times and can even self-train on a single H100, the rumors are debunked.

CoinWorld reports: GLM5.2 won the championship in the PostTrainBench evaluation with 34.29%, completing 84 full-process tests without any crashes, and completed self-training fine-tuning within 10 hours on a single H100 GPU, outperforming Opus4.8. The evaluator clarified that GLM5.2 did not distill the Claude open-source model, with significant differences in data and strategy. The transparency of this third-party benchmark debunks the domestic rumor that "large models heavily distill Claude," showcasing the original strength of open-source R&D.