Cursor 承認 Composer 2 底層是中國 AI 模型 Kimi K2.5：開發工具的透明度危機 | Cursor Admits Composer 2 Is Built on Chinese AI Kimi K2.5: A Developer Trust Crisis
By Kit 小克 | AI Tool Observer | 2026-03-29
🇹🇼 Cursor 承認 Composer 2 底層是中國 AI 模型 Kimi K2.5：開發工具的透明度危機
2026 年 3 月 22 日，AI 編程工具 Cursor 發布了 Composer 2，宣稱這是「前沿等級的編碼智慧」，是自家研發的成果。幾小時之內，一位名叫 Fynn 的開發者用封包分析工具攔截了 API 流量，發現底層模型 ID 是：accounts/anysphere/models/kimi-k2p5-rl-0317。
沒錯——Cursor 沒有告訴用戶，Composer 2 的核心是來自中國新創公司 Moonshot AI（月之暗面）的開源模型 Kimi K2.5，而 Moonshot AI 的背後投資方包括阿里巴巴與騰訊。
一個 API 呼叫，拆穿了行銷話術
Fynn 的方法並不複雜：他用 Charles Proxy 或類似工具監測 Cursor 送出的 HTTP 請求，在 payload 裡直接看到了模型名稱。Cursor 聯合創辦人事後在 Reddit 承認，Composer 2 確實是在 Kimi K2.5 上進行 fine-tuning 的產物，並非從頭訓練的自研模型。
- 模型基座：Kimi K2.5（Moonshot AI，Apache 2.0 授權）
- Cursor 的工作：針對編碼任務進行微調（fine-tuning）
- 發布時的說明：幾乎為零，無任何主動揭露
這不是 Cursor 第一次藏著掖著
更讓社群不安的是，這已經是 Cursor 第二次被發現悄悄使用中國 AI 技術。Composer 1 曾被發現底層使用了 DeepSeek 的 tokenizer，同樣沒有對外說明。兩次事件連在一起，讓不少開發者開始質疑：Cursor 對底層技術選型的保密，究竟是商業保護，還是刻意迴避地緣政治敏感性？
為什麼這件事值得你在意？
有人說：「反正開源模型大家都可以用，有什麼問題？」但問題不在模型本身，而在於：
- 你的程式碼被送到哪裡？ Cursor 的 cloud 模式下，用戶的程式碼會上傳至 Anysphere 的伺服器。底層模型的來源影響的是信任鏈與資料主權。
- 企業合規風險： 部分企業的資安政策明確禁止程式碼流向中國技術棧。若 Cursor 未主動揭露，IT 部門根本無從稽核。
- 行銷與現實的落差： 「前沿等級的編碼智慧」這種措辭，讓用戶以為 Cursor 有自己的研究實力——但其實只是在做 fine-tuning。
開發者現在怎麼辦？
短期不需要立刻換工具。Kimi K2.5 是合法的開源模型，fine-tuning 本身也是業界常見做法。但你應該：
- 在 Cursor 設定中確認使用的是哪個模型，以及是否為 cloud 模式
- 敏感專案考慮切換到 本地部署模式（如 Continue.dev + 本地模型）
- 企業用戶向 Cursor 要求提供完整的 model card 與資料處理說明
透明度不是選配，是信任的基礎。這次事件最大的教訓不是「不該用中國模型」，而是：工具商有義務告訴你，你的程式碼在跟誰說話。
好不好用，試了才知道。
🇺🇸 Cursor Admits Composer 2 Is Built on Chinese AI Kimi K2.5: A Developer Trust Crisis
On March 22, 2026, AI coding tool Cursor launched Composer 2, marketing it as "frontier-level coding intelligence" built in-house. Within hours, a developer named Fynn intercepted the API traffic and found the underlying model ID: accounts/anysphere/models/kimi-k2p5-rl-0317.
That is Kimi K2.5 — an open-source model from Chinese startup Moonshot AI, backed by Alibaba and Tencent. Cursor had not disclosed this to users at launch.
One API Call Broke the Narrative
Fynn's method was straightforward: he used a packet-inspection proxy to monitor outgoing HTTPS requests from Cursor and spotted the model name directly in the payload. Cursor's co-founder later confirmed on Reddit that Composer 2 is indeed fine-tuned from Kimi K2.5 — not trained from scratch by Anysphere's research team.
- Base model: Kimi K2.5 (Moonshot AI, Apache 2.0 license)
- Cursor's contribution: Fine-tuning on coding tasks
- Disclosure at launch: None — zero proactive communication
This Is the Second Time
What makes the community more uneasy is that this is Cursor's second undisclosed use of Chinese AI technology. Composer 1 was previously found to use DeepSeek's tokenizer under the hood, also without announcement. Together, these incidents raise a fair question: is Cursor's secrecy about model sourcing commercial confidentiality, or deliberate avoidance of geopolitical scrutiny?
Why This Actually Matters
Some developers shrug: "It's open-source, anyone can use it." But the issue isn't the model itself — it's about:
- Where does your code go? In Cursor's cloud mode, your code is uploaded to Anysphere's servers. The origin of the underlying model matters for the trust chain and data sovereignty.
- Enterprise compliance risk: Many company security policies explicitly prohibit code flowing through Chinese technology stacks. Without disclosure, IT teams cannot audit this.
- Marketing vs. reality gap: Phrases like "frontier-level coding intelligence" imply proprietary research capability — not "we fine-tuned someone else's open-source model."
What Should Developers Do?
There's no need to panic or immediately switch tools. Kimi K2.5 is a legitimate open-source model, and fine-tuning is standard industry practice. But here's a practical checklist:
- Check your Cursor settings to confirm which model is active and whether you're in cloud mode
- For sensitive projects, consider switching to local deployment (e.g., Continue.dev + local models)
- Enterprise users should formally request a complete model card and data-handling documentation from Cursor
Transparency isn't a nice-to-have — it's the foundation of trust. The lesson here isn't "don't use Chinese models." It's this: tool vendors have an obligation to tell you who your code is talking to.
You won't know until you try it — but first, you deserve to know what you're trying.
Sources / 資料來源
- Cursor Admits Its New Coding Model Was Built on Top of Moonshot AI's Kimi — TechCrunch
- Cursor's Composer 2 Was Secretly Built on a Chinese AI Model — VentureBeat
- Cursor Faces Backlash After Revealing Its Coding Model Was Built on Kimi K2.5 — Benzinga
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
張貼留言