Kimi K3 蒸餾爭議：白宮指控月之暗面竊取 Claude 技術 | Kimi K3 Distillation Row: White House Accuses Moonshot

7月 23, 2026

By Kit 小克 | AI Tool Observer | 2026-07-23

🇹🇼 Kimi K3 蒸餾爭議：白宮指控月之暗面竊取 Claude 技術

白宮科技政策辦公室主任 Michael Kratsios 於 7 月 22 日公開指控中國新創 月之暗面（Moonshot AI）透過大規模「蒸餾」（distillation）手法竊取 Anthropic 旗艦模型 Claude Fable 的能力，用來訓練其最新開源模型 Kimi K3。這起爭議不僅牽動中美 AI 技術戰的敏感神經，也讓「蒸餾」這個原本純屬技術名詞的詞彙，瞬間成為地緣政治攻防的焦點。

什麼是「蒸餾」？合法技術還是竊取行為？

模型蒸餾（model distillation）是業界常見手法：讓體積較小的模型模仿大型模型的輸出，藉此壓縮出效能相近但成本更低的版本。Kratsios 也坦承這項技術本身合法，但他強調 Moonshot 的做法「規模龐大且刻意隱匿」，形容其為「大規模、隱密的產業級蒸餾，目的是竊取受保護的美國技術」。

證據一：340 萬次偽造帳號對話

Anthropic 先前已指控 Moonshot 透過詐欺帳號，累積產生超過 340 萬次 Claude 對話紀錄，藉此萃取推理、程式撰寫、工具使用與電腦視覺等核心能力，繞過官方存取限制。更戲劇性的是，部分使用者回報 Kimi K3 在對話中曾直接自稱是「Claude」，被視為蒸餾痕跡的直接證據。

證據二：疑似透過泰國取得禁售 Nvidia 晶片

Kratsios 進一步指控 Moonshot 透過泰國取得原本禁止對中國出口的 Nvidia GB300（Blackwell 架構）伺服器，用於訓練其模型。財政部長 Scott Bessent 隨後表態，不排除祭出制裁與出口管制黑名單。

為何這件事現在爆發？

Kimi K3 是一款 2.8 兆參數的開源模型，於 7 月 16 日發布後迅速在 Frontend Code Arena 拿下對 Claude Fable 5 高達 76% 的勝率，成為史上最大規模的開源權重釋出，甚至一度導致美國 AI 相關股票下挫，投資人擔憂其對 OpenAI、Anthropic 等公司構成實質競爭威脅。此刻美方的指控，被外界解讀為技術實力落後之外的另一種反制手段。

對開發者與產業的意義

開源模型信任問題：若蒸餾指控屬實，開發者在選用開源模型時需更留意其訓練資料來源與授權合規性。
晶片出口管制可能收緊：若制裁落地，將進一步影響中國 AI 公司取得高階算力的管道。
中美 AI 競賽持續白熱化：無論指控真假，這都顯示美中在前沿模型能力上的差距正快速縮小。

目前 Moonshot AI 尚未對指控做出正式回應。這起事件也提醒我們，開源模型的「免費」背後，可能牽涉更複雜的技術與地緣政治成本。

好不好用，試了才知道。

🇺🇸 Kimi K3 Distillation Row: White House Accuses Moonshot

White House Office of Science and Technology Policy director Michael Kratsios publicly accused Chinese startup Moonshot AI on July 22 of using large-scale distillation to extract capabilities from Anthropic flagship model Claude Fable in order to train its new open-weight model, Kimi K3. The allegation has turned a routine machine-learning technique into the latest flashpoint in the US-China AI rivalry.

What Is Model Distillation, and Is It Theft?

Model distillation, training a smaller model to mimic a larger one output, is a standard, legal technique used across the industry to build cheaper, faster models. Kratsios acknowledged this, but argued Moonshot crossed a line, describing its approach as large-scale, covert industrial distillation designed to steal protected U.S. technology.

Exhibit A: 3.4 Million Fraudulent Claude Exchanges

Anthropic had previously alleged that Moonshot generated more than 3.4 million Claude conversations through fraudulent accounts to extract reasoning, coding, tool-use, and computer vision capabilities while bypassing access restrictions. Adding fuel to the claim, some users reported that Kimi K3 identified itself as Claude mid-conversation, seen by critics as a smoking gun for distillation.

Exhibit B: Banned Nvidia Chips via Thailand

Kratsios also alleged Moonshot obtained export-restricted Nvidia GB300 Blackwell servers through Thailand to train its models. Treasury Secretary Scott Bessent followed up by floating sanctions and export-control blacklisting as possible responses.

Why Now?

Kimi K3 is a 2.8-trillion-parameter open-weight model released July 16 that quickly posted a 76% win rate over Claude Fable 5 on the Frontend Code Arena, the largest open-weight release to date, and one that briefly rattled US AI stocks as investors worried about real competitive pressure on OpenAI and Anthropic. The timing of the accusation suggests it is as much a competitive response as a legal one.

What This Means for Developers

Open-weight trust questions: developers picking open models should pay closer attention to training-data provenance and licensing compliance.
Tighter chip export controls could follow if sanctions materialize, further squeezing Chinese labs access to frontier compute.
The US-China AI gap keeps narrowing, true or not, the accusation itself signals how seriously Washington is taking Chinese open-weight progress.

Moonshot AI has not yet issued a formal response to the allegations. The episode is a reminder that free open-weight models can carry hidden technical and geopolitical costs.

好不好用，試了才知道 — the only way to know if it is worth it is to try it yourself.

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

搜尋此網誌

AI小貼士