deepseek-v4.1IMAGE

DeepSeek V4.1 Flash 1017 retested on OfficeBench's 300 tasks: more accurate, faster, cheaper

The author reran the public OfficeBench 300-task set against the previous build and reports build 1017 gains accuracy, finishes faster, and uses fewer tokens, lowering cost.

林高行

@takayuki884sanArchived 11 days agoOriginally posted 2026.10.20

DeepSeek V4.1 Flash 1017 retested on OfficeBench's 300 tasks: more accurate, faster, cheaper (image 1 of 2)

DeepSeek V4.1 Flash 1017 retested on OfficeBench's 300 tasks: more accurate, faster, cheaper (image 2 of 2)

林高行 @takayuki884sanOriginally posted 2026.10.20

Open original on X

DeepSeek V4.1 Flash:1017が公開されたので、OfficeBench（公開の学術ベンチマーク・公式300課題）で測り直しました。 1つ前の同モデル、ビルドと比べると、推論能力が上がるにつれて、 1. 正答率の上昇 2. 所要時間が短く 3. 2のせいで使用トークンが少なくなり、コスト改善

x.com/takayuki884san/status/2112377896102928895Archived 2026.10.20 12:03