DeepSeek V4.1 Flash 1017 发布：MIT 开源权重与跑分变化
官方发布信息汇总：Terminal-Bench 4.0 从 31.2 升至 42.3、HLE 36.8→42.6、programBench 20.3→29.1、DeepSWE 77.0 超过 Opus 5 与 GPT 5.6 Sol，API 价格不变，网页与 App 已上线。
DEEPSEEK JUST DROPPED A NEW OPEN MODEL😳 DeepSeek V4.1 Flash 1017 is live today Open weights. MIT license. free to download What changed -terminal Bench 4.0 went from 31.2 to 42.3 -HLE went from 36.8 to 42.6 -programBench went from 20.3 to 29.1 -deepSWE hit 77.0, ahead of Opus 5 and GPT 5.6 Sol on their own table -faster runs and shorter answers, so tasks cost less Same API price as before The DeepSeek web app and phone app already use it How to try it 1. Open http://chat.deepseek.com and just start chatting 2. On the API, call the model name deepseek-flash 3. Want it local? Grab the weights on Hugging Face Model card https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017 Changelog https://api-docs.deepseek.com/updates
这条是引用着另一条写的：x.com/alizeh_ai/status/2111290983312933369