Startups including Hirundo and Lazarus AI are editing the weights in Chinese open-weight models to remove censorship, state propaganda, and security weaknesses for wary Western companies, according to Rest of World. Hirundo’s own tests found Alibaba’s Qwen models won’t acknowledge the Tiananmen Square crackdown or Xinjiang’s labor camps, and CEO Ben Luria said its edits bring the models in line with ChatGPT and Claude. Lazarus AI stripped Qwen’s censorship by retraining the models on Gemini Flash answers to prompts built from leaked censorship guidelines. Lazarus says its ReAligned models run at about a dozen US companies, but some government clients still refuse Chinese models. Stanford DigiChina Project editor-in-chief Graham Webster said most users don’t care how the models are aligned.
