deepseek-v4.1POST
官方表对比：V4.1 Flash 编程 74.2 略胜 Opus 5，HLE 差 20 分
按 DeepSeek 官方表，V4.1 Flash 在 DeepSWE 上 74.2 对 Opus 5 的 74.0，Humanity's Last Exam 则 36.8 对 56.3，作者总结编程追平、硬推理仍差约 20 分。
Uzair Afridi
@uzair__afridi11 天前收录原推发布于 2026.10.20
Uzair Afridi @uzair__afridi原推发布于 2026.10.20
Same model, two stories. DeepSeek V4.1-Flash, per DeepSeek's own table: DeepSWE: 74.2 vs Opus 5's 74.0 Humanity's Last Exam: 36.8 vs 56.3 Open weights caught up on coding. Hard reasoning is still a 20-point gap. Which matters for your workload?
x.com/uzair__afridi/status/2112242382385336759采集于 2026.10.20 12:05