近期中國頂尖AI實驗室（如智譜、月之暗面和阿里）相繼推出強大的開源模型（GLM 5.2、Kimi K3、Qwen 3.8），效能逼近西方頂尖閉源模型且在智慧體程式碼編寫上表現優異。這引起了美國矽谷與政界的廣泛關注，甚至面臨美國政府關於技術蒸餾的指控與潛在制裁。

相較於西方實驗室（如Anthropic和OpenAI）因安全疑慮及美國政府的出口管制，逐漸將 Mythos 或 GPT 等尖端模型實施限制或延遲發布，中國公司則加倍投資開源領域。這讓開發者可以下載權重、在本地運行並高度自訂，使美中AI競爭逐漸演變為「開源」對抗「閉源」的路線之爭。

這些中國開源模型已成為西方初創企業與研究人員的實用替代方案，特別是在網路安全等閉源模型常因安全限制而拒絕協助的領域。雖然Kimi K3等模型可能消耗較多Token，但它們的出現打破了「只有無限制資金和算力才能提升AI效能」的假設，對西方閉源巨頭的商業模式構成直接挑戰。

Recently, leading Chinese AI labs (such as Z.ai, Moonshot AI, and Alibaba) have released powerful open-source models (GLM 5.2, Kimi K3, Qwen 3.8) that perform nearly as well as top Western models and excel at agentic coding. This has triggered intense interest and concern in Silicon Valley and Washington, leading to allegations of distilling proprietary U.S. technology and threats of sanctions.

While Western labs like Anthropic and OpenAI are increasingly locking down their frontier models due to safety concerns and U.S. government export controls, Chinese firms are doubling down on open weights. This approach allows users to run models locally and customize them freely, turning the geopolitical rivalry into a clear debate between open-source and closed-source strategies.

These Chinese open-source models are becoming practical commercial alternatives for Western startups and researchers, particularly in areas like cybersecurity where proprietary Western models refuse to cooperate due to strict safety guardrails. Although models like K3 might be token-hungry, their capability challenges the core assumption that infinite compute funding is required to build frontier AI.