跳到主要內容

DeepSeek V4.1 Flash評測:MIT開源,編碼跑分互有輸贏 | DeepSeek V4.1 Flash Review: MIT Weights, Mixed Benchmarks

By Kit 小克 | AI Tool Observer | 2026-10-18

🇹🇼 DeepSeek V4.1 Flash評測:MIT開源,編碼跑分互有輸贏

DeepSeek V4.1 Flash 在10月17日無預警放出新版本,代號 1017,直接把權重丟上 Hugging Face 和 ModelScope,用的是完全沒有商用限制的 MIT 授權。這次更新主打三件事:編碼跑分、1M 超長上下文、以及第一次對外開放自家的 Search 與 Fetch API。對天天用 AI 寫程式的人來說,這次更新比上一代 Flash 進步不小,但也不是全面碾壓對手。

跑分:編碼贏一半,Agent任務還是輸

先說好消息:在 DeepSWE v1.1(專門測真實世界修 bug 能力的跑分)上,DeepSeek V4.1 Flash 拿到 77.0 分,不只贏過自家上一版的 74.2,也小勝 Anthropic 的 Claude Opus 5(74.0)和 OpenAI 的 GPT-5.6 Sol(73.0)。

但換到 Terminal-Bench 4.0(測多步驟終端機操作,更接近真實 Agent 工作流),V4.1 Flash 只拿到 42.3,雖然比上一版的 31.2 大幅進步,卻還是輸給 Claude Opus 5 的 51.8。換句話說,這次更新讓 DeepSeek 在「單次編碼任務」上追平甚至超車一線模型,但在「長時間自主操作」上還有明顯差距。不要被跑分截圖騙了,實際拿來跑 agent 工作流前,自己測一輪比較保險。

價格:比對手便宜一個量級

  • 輸入(未命中快取):離峰每百萬 token $0.15,尖峰 $0.30
  • 輸出:離峰每百萬 token $0.60,尖峰 $1.20
  • 命中快取的輸入甚至低到每百萬 token $0.003~$0.006

對比 Claude Opus 5、GPT-5.6 這類一線模型,DeepSeek V4.1 Flash 的價格大概只是零頭,對於需要跑大量重複性編碼任務、預算有限的團隊會是划算選項。

Search 與 Fetch API 首次開放

這次更新同時開放了 DeepSeek 自家的搜尋與網頁擷取 API,每千次成功呼叫收費 $0.75,單帳號限速 5 QPS。更實際的是,透過 DeepSeek Harness、Claude Code、Codex 這些工具呼叫時,10月31日 UTC 16:00 前都免費,等於給開發者一段時間先試用再決定要不要正式接入。

值得換嗎

如果你的工作流主要是單次程式碼生成或修 bug,這次更新加上價格優勢,確實值得排進測試清單;但如果你依賴的是長時間、多步驟的自主 agent 操作,Terminal-Bench 的落差提醒你,現階段還是該讓一線模型扛主力。好不好用,試了才知道。


🇺🇸 DeepSeek V4.1 Flash Review: MIT Weights, Mixed Benchmarks

DeepSeek V4.1 Flash dropped without warning on October 17, tagged build "1017," with weights posted straight to Hugging Face and ModelScope under a fully permissive MIT license. The update leads with three things: coding benchmarks, a 1M-token context window, and the first public opening of DeepSeek's own Search and Fetch APIs. If you write code with AI daily, this is a real step up from the previous Flash — but it is not a clean sweep over the competition.

Benchmarks: Wins Coding, Still Loses Agent Tasks

The good news first: on DeepSWE v1.1, a benchmark built around real-world bug fixing, DeepSeek V4.1 Flash scores 77.0 — ahead of its own predecessor's 74.2, and narrowly past Anthropic's Claude Opus 5 (74.0) and OpenAI's GPT-5.6 Sol (73.0).

Switch to Terminal-Bench 4.0, which measures multi-step terminal operations closer to real agent workflows, and V4.1 Flash drops to 42.3. That is a big jump from the previous version's 31.2, but still well behind Claude Opus 5's 51.8. In short: this release closes the gap on single-shot coding tasks, even edging past frontier models, but long-horizon autonomous work is still a clear weak spot. Do not trust the benchmark screenshot alone — run your own test before betting an agent pipeline on it.

Pricing: An Order of Magnitude Cheaper

  • Input (cache miss): $0.15 per million tokens off-peak, $0.30 peak
  • Output: $0.60 per million tokens off-peak, $1.20 peak
  • Cache-hit input drops as low as $0.003–$0.006 per million tokens

Compared to frontier models like Claude Opus 5 or GPT-5.6, DeepSeek V4.1 Flash costs a fraction of the price — a solid option for teams running high-volume, repetitive coding tasks on a tight budget.

Search and Fetch APIs Go Public

This release also opens DeepSeek's internal search and web-fetch APIs for the first time, priced at $0.75 per 1,000 successful calls, capped at 5 QPS per account. More practically, calls made through DeepSeek Harness, Claude Code, or Codex as server-side tools are free until October 31 at 16:00 UTC, giving developers a window to test before billing kicks in.

Is It Worth Switching?

If your workflow is mostly single-shot code generation or bug fixing, this update plus the price advantage makes it worth adding to your test list. But if you lean on long, multi-step autonomous agent runs, the Terminal-Bench gap is a reminder that frontier models still carry the heavy lifting for now. 好不好用,試了才知道 — only testing it yourself will tell.

Sources / 資料來源

延伸閱讀 / Related Articles


AI 工具觀察站 — 每日精選 AI Agent 與工具趨勢
AI Tool Observer — Daily curated AI Agent & tool trends

留言

這個網誌中的熱門文章

Google Ironwood TPU v7 推理專用晶片解析:效能追平 NVIDIA、成本低 44%,AI 晶片戰爭正式開打 | Google Ironwood TPU v7 Explained: Matching NVIDIA Performance at 44% Lower Cost — The AI Chip War Heats Up

Claude Code 實測:AI 幫你寫程式到底行不行? | Claude Code Review: Can AI Really Code for You?

Cursor vs GitHub Copilot vs Claude Code:AI 程式助手大比拼 | AI Coding Assistants Compared: Cursor vs GitHub Copilot vs Claude Code