西方技术初创企业如 Hirundo 和 Lazarus AI 正利用机器遗忘和权重编辑技术，剥离中国开源权重人工智能模型（如阿里巴巴的 Qwen 和 DeepSeek）中国家强制要求的审查机制，并修补安全漏洞。
Western Startups Modify Chinese Open-Weight AI Models to Remove Censorship and Bolster Security Western technology startups such as Hirundo and Lazarus AI are using machine unlearning and weight-editing techniques to strip state-mandated censorship and patch security flaws in Chinese open-weight artificial intelligence models like Alibaba's Qwen and DeepSeek.
尽管这些高性价比模型已在全球获得广泛采用，但西方企业和政府客户仍担忧政治偏见、网络漏洞及监管禁令。通过使用符合西方标准的数据重新对齐模型，这些初创企业旨在让中国开源系统适配企业级部署，但专家指出，制定通用的对齐标准仍是一个持续面临的挑战。
While these cost-effective models have gained broad global adoption, Western corporate and government clients face concerns over political bias, cyber vulnerabilities, and regulatory prohibitions. By realigning the models using Western-aligned data, these startups aim to make Chinese open-source systems viable for enterprise deployment, though experts note that defining universal alignment standards remains an ongoing challenge.