V4.1 Flash 仅 8B–16B 激活参数，DeepSWE 77% 超闭源

Frank trading @Franktradinglog原推发布于 2026.10.18

在 X 上打开原推

As I've said before, technological diffusion is an irreversible process, and China has consistently outpaced the US when it comes to driving down costs. The technological moat enjoyed by closed-source frontier AI labs simply cannot last forever. The latest release of DeepSeek-V4.1-Flash-1017 is yet another reminder of this reality. With only 8B–16B activated parameters, the model reportedly achieved a 77% Pass@1 score on DeepSWE v1.1, outperforming several leading proprietary models on this particular benchmark at a fraction of the cost. This is a fundamental shift. Open-source models are no longer just "good enough" alternatives to proprietary models. They are increasingly capable of matching—and potentially surpassing—the most advanced closed-source models in specific domains, while dramatically undercutting their economics. And this process is only accelerating. Eventually, we'll see highly capable AI models running on a single RTX 5090, or even an ordinary laptop. As intelligence becomes increasingly commoditized and the cost of inference continues to collapse, the scarcity premium embedded in today's AI valuations will become harder and harder to justify. At that point, continuing to trade the so-called "bottleneck" narrative as if compute scarcity were a permanent structural advantage would be absurd. The AI trade as we know it is coming to an end.

这条是引用着另一条写的：x.com/franktradinglog/status/2053106345658966244