AI 导读
5/ 注: - 这 200 个案例是合成的,均匀分布在 30 种任务类型上 - LLM 的推理功能已关闭,但 GLM 5.3 Flash 除外,它需要推理且以低 effort 运行 - DeepSeek 和 GLM 使用了默认的提供商路由,因此它们的尾部延迟反映了各提供商的混合情况
正文
5/ Note:
- the 200 cases are synthetic, evenly spread across 30 task types
- reasoning was off for LLMs, except GLM 5.3 Flash, which requires reasoning and ran at low effort
- DeepSeek and GLM used default provider routing, so their tail latency reflects the mix of providers
来源:@OpenRouter · x.com