deepseek-v4.1POST
Terminal-Bench 4.0 得 42.3 分、每任务 0.20 美元的成本账
作者认为 MIT 开源权重在同等跑分下成本低 10–20 倍，继续为闭源前沿 API 付溢价是在烧钱。
Mo Ayob
@Mossiah12 天前收录原推发布于 2026.10.18
Mo Ayob @Mossiah原推发布于 2026.10.18
Scoring 42.3 on Terminal-Bench 4.0 at $0.20 per task makes running agent loops on closed frontier APIs look reckless. When MIT open weights hit these benchmarks at 10x to 20x lower cost, paying proprietary API markups is just burning runway. https://x.com/deepseek_ai/status/2111375640717033777
这条是引用着另一条写的：x.com/deepseek_ai/status/2111375640717033777
x.com/mossiah/status/2111742830905360609采集于 2026.10.18 17:40