deepseek-v4.1POST
Why OpenCode call volume and Agent Arena rank measure different things
The author argues OpenCode's 24h call volume reflects real engineering usage where V4.1 Flash leads on low latency, while Agent Arena measures quality ceiling, and suggests comparing pass@k against end-to-end latency on the same projects.
Hale
@YunTechXArchived 2 days agoOriginally posted 2026.10.26
Hale @YunTechXOriginally posted 2026.10.26
不矛盾，两边在量不同的东西。OpenCode 24h 调用量量的是谁在被工程流程实际调用--DeepSeek V4.1 Flash 推理延迟低、部署面广，工程循环里跑得最多，调用量自然断层。Agent Arena 量的是给定任务的质量上限，所以 Kimi K3.1 Pro 能拿开源第一。一句话：Flash 赢的是「能跑、跑得快」，K3.1 Pro 赢的是「跑得好」。要判断该用谁，拿同一批项目跑 SWE-verify 或 LiveCodeBench，看 pass@k 和端到端延迟的 tradeoff，别拿一个榜单解释另一个。
x.com/yuntechx/status/2114711940408779083Archived 2026.10.27 12:10