上手吐槽 V4.1-Flash-1017：幻觉、速度感知与 harness 依赖
实测反馈五点：幻觉仍需第二模型复核；省 token 不等于更快，部分测试反而更慢；质量强依赖 harness；多文件重构需更强规划；Search+Fetch 11 月 1 日起按次收费。
a few things i don’t like about DeepSeek V4.1-Flash-1017 so far: 1. hallucinations are still the biggest issue. i wouldn’t trust it to approve its own architectural changes without a second model checking the work. 2. “faster” is a bit misleading. the model can finish tasks with fewer tokens, but that doesn’t always mean higher tokens/sec, and in some tests 1017 actually feels slower. 3. the quality still depends heavily on the harness. DeepSeek specifically optimized 1017 for a few environments, which means the same model may not feel equally good everywhere. 4. complex multi-file refactors are still where i’m most cautious. it can implement fast, but once the task crosses architectural boundaries i want a stronger planner involved. 5. the new Search + Fetch stack is useful, but it’s another dependency that will start charging per call on November 1. cheap, yes: free, no. 1017 is clearly better. but it still works best for me as a very fast worker, not the final authority.
这条是引用着另一条写的：「DeepSeek V4.1-Flash-1017 上手观察」