Industry

Kimi 博客数据与 Terminal-Bench 4.0 结果不一致引发质疑

Why we picked it

有人发现 Kimi 博客表格里 42.3 这个分其实是 DeepSeek 发布的数据,和图表脚注对不上,基准确实容易被各家包装,看看细节。

X 用户 teortaxesTex 指出 Kimi 博客与 Terminal-Bench 4.0 成绩存在不一致。博客表格中 Kimi 1017 模型的分数为 42.3,该数字实际来自 DeepSeek 的发布内容。同时图表脚注写明,除 DeepSeek 外所有模型的成绩均来自 tbench.ai,成本数据来自 Artificial Analysis。这一矛盾让博客基准数据的可信度受到质疑。

Original · Teortaxes

What's sillier is that Kimi blog is inconsistent with TB 4.0 results. In the table they show 42.3 for 1017, which is from DS's release. In the chart, > Performance and cost data for all models are sourced from tbench ai, except for DeepSeek, whose data come from @ArtificialAnlys https://t.co/mXyXFrwOEy