A crop of startups is rewriting the weights of Chinese open models such as DeepSeek and Qwen, removing censorship and security risks for Western firms worried about both.

Chinese open models like DeepSeek are cheap and strong, and Western businesses want them but fear them, so a new business has appeared: editing the weights.
The most concrete finding comes from security firm CrowdStrike: prompts to DeepSeek's R1 on politically sensitive topics were up to 50% more likely to return code with severe security vulnerabilities.
Another firm, Lazarus AI, realigned Alibaba's Qwen using answers from Google's Gemini Flash, and says the models could then name sensitive activists not present in their training material, suggesting the knowledge had been there but suppressed.
That a few weight edits can swap a model's stance unsettles me a little: what does the model actually know?



Curated from high-quality sources, with concise summaries and key takeaways.