13:I[28705,["5004","static/chunks/70e0d97a-78d972e0bdc586fd.js","277","static/chunks/54a60aa6-04fcfe43859edf73.js","9352","static/chunks/9352-be4f984da86eb3d1.js","9641","static/chunks/9641-b36240abade64990.js","6811","static/chunks/6811-a83d713117d26386.js","6874","static/chunks/6874-c4e6f31cb2e95bcf.js","2631","static/chunks/2631-9f01cb1e5bd0a86c.js","7486","static/chunks/7486-6f0b189a58f0663d.js","6719","static/chunks/6719-b5f6a8a1ec6b8016.js","1834","static/chunks/1834-00b155d9aeed73d1.js","5564","static/chunks/5564-af00d10d47414246.js","1438","static/chunks/1438-e31bfa8da7355a38.js","546","static/chunks/546-88e96de68aff5232.js","6788","static/chunks/6788-9db81b79e48ef15c.js","3553","static/chunks/3553-cf97d49fed029d0e.js","206","static/chunks/206-006ab51bcb84ff25.js","7283","static/chunks/7283-1eb35b5617426157.js","8617","static/chunks/app/card/%5Bid%5D/page-92b890d96ab49714.js"],"CardDetailContent"] 14:T704,

昨天深夜 DeepSeek V3.1 發布了，今天看了一下有幾個值得注意的更新：

混合推理架構：一個模型支援「思考模式」與「非思考模式」，用戶可以在官方 App 裡直接切換。這其實跟最近大家在討論的「多階段 reasoning」概念有點呼應。
思考效率提升：跟之前的 R1-0528 比，V3.1 在思考模式下 token 數減少 20%-50%，但表現不掉，代表他們做了 chain-of-thought 壓縮。這對落地應用蠻關鍵，因為推理成本下降。
Agent 能力：特別提到 post-training 強化了工具使用與 Agent 任務表現，像 SWE、Terminal-Bench、browsecomp、HLE 這些 benchmark 都有顯著提升，意味著他們真的在針對「實際 agent 場景」調校。

V3.1 Base Open-source weights:

DeepSeek V3.1 發布了

留言區

請先登入...