A growing number of startups are editing powerful Chinese open-weight AI models to strip out censorship and harden their security for Western businesses, according to a report by Rest of World. The firms, including Israel-based Hirundo and Boston-based Lazarus AI, use techniques such as machine unlearning and direct weight edits to realign models from DeepSeek, Alibaba's Qwen, Moonshot, Z.AI and Tencent.
The trend took off after DeepSeek released its first open-weight large language model last year. Chinese models have won over users, including American businesses, with lower costs, strong capabilities and customization options. But some Western companies worry the models are too closely aligned with the Chinese Communist Party and may not be as safe as closed systems.
How machine unlearning realigns Chinese AI models
Hirundo co-founder and CEO Ben Luria described the process as "neurosurgery" — a surgical intervention that targets values, geopolitical perceptions, censorship and state-aligned bias. His team edits the weights of open models to make them more secure against cyberattacks, improve fairness and reduce dangerous knowledge such as information about bioweapons.
Lazarus AI took a different route with Alibaba's Qwen models. It identified banned topics from leaked censorship guidelines, generated prompts around those topics, and used answers from Google's Gemini Flash to realign the models. After the intervention, the models could produce information that was not in their training materials, such as names of pro-democracy activists, suggesting the knowledge had been learned but suppressed.
- Hirundo edits model weights to realign values and remove state-aligned bias
- Lazarus AI used leaked censorship guidelines and Gemini Flash answers to retrain Qwen models
