Artificial Analysis· @ArtificialAnlys · X·· 2 小时前AI 评分38
AI 导读
GLM-5.3 在覆盖范围内拥有最多的端点，共 26 个。评分范围从 67.2 到 99.3，其中 16 个在统计上与参考值一致，10 个端点评分低于参考值。最大的三个差距来自输出 token 上限和推理循环截断了 HLE-250 的答案；量化并非原因，六个 FP4 端点中有五个与 FP8 参考值一致。
正文
GLM-5.3 has the most endpoints in coverage at 26. Scores range from 67.2 to 99.3, with 16 statistically in line with the reference and 10 endpoints scoring below. The three largest gaps come from output token caps and reasoning loops that cut off HLE-250 answers; quantization is not a cause, with five of six FP4 endpoints in line with the FP8 reference
来源：Artificial Analysis · x.com