DeepSeek 發布 V4.1-Flash-1017,權重以 MIT 授權開放,API 價格不變DeepSeek releases V4.1-Flash-1017 with MIT-licensed weights at unchanged API prices
- 事件
- DeepSeek 於十月十七日發布 V4.1-Flash-1017,權重以 MIT 授權在 Hugging Face 公開,支援一百萬 token 上下文;API 每百萬輸入零點一五美元、輸出零點六美元(離峰)。DeepSeek released V4.1-Flash-1017 on 17 October. The weights are public under the MIT licence on Hugging Face and ModelScope with a one-million-token context, and API pricing is unchanged at 0.15 US dollars per million input tokens and 0.60 per million output tokens off-peak.
- 背後
- 其自行公布的測試中,Terminal-Bench 4.0 由三十一點二升至四十二點三,DeepSWE 一點一由七十四點二升至七十七,但同表顯示 Claude Opus 5 在知識類測試仍領先。DeepSeek's own card reports Terminal-Bench 4.0 rising from 31.2 to 42.3 and DeepSWE 1.1 from 74.2 to 77.0, while the same table still shows Claude Opus 5 ahead on knowledge-heavy evaluations. None of the figures had been independently reproduced.
我們認為,開放權重加上不變價格,把模型採購由單價競賽改為總成本競賽;企業真正的門檻不在取得模型,而在能否自行驗證它在自己流程上的表現。We believe open weights at unchanged prices turn model procurement from a price-per-token contest into a total-cost contest. The real constraint for an enterprise is not access to the model but its ability to verify how it behaves inside its own workflows.
- 對日常工作的影響
- IT 需自建一組固定測試題與評分口徑,採購談判改以每完成一個任務的成本計價,財務毋須再按月鎖定單一供應商額度。IT must build a fixed set of test prompts with a scoring rubric, procurement should negotiate on cost per completed task rather than per token, and finance no longer needs to lock one vendor's monthly allocation.
- 如何改善
- 本週整理十個內部真實任務作測試集,用同一口徑比較開放權重與現用模型的完成率與每任務成本。This week, assemble ten real internal tasks as a test set and compare the open-weight model with your current vendor on completion rate and cost per task under one scoring method.
