Two Western companies have built businesses on the same premise: Chinese open-weight models are cheap and capable, but enterprises want the values layer removed before they deploy them.
Hirundo, founded by Israeli entrepreneur Ben Luria, edits weights directly. The company describes the work as improving security, fairness and behaviour on topics Beijing treats as sensitive. Boston-based Lazarus AI takes a different route, removing what it calls mandatory censorship from Alibaba’s Qwen line.
How the realignment is done
Lazarus built prompts from leaked censorship guidelines, then used answers from Google’s Gemini Flash as a signal to shift the model. Eric Hartford, formerly the company’s chief scientist, said the edited models went on to produce information absent from their training material, such as names of pro-democracy activists.
Roughly a dozen US enterprises have taken the realigned variants, the company says. Hirundo reports its own tests showing Qwen declines to engage with certain historical subjects entirely.
The safety case rests on outside research. A CrowdStrike analysis of DeepSeek’s R1 found prompts involving politically sensitive topics were up to 50 percent more likely to return code with severe vulnerabilities.
What editing does not settle
Alignment can run deeper than a targeted intervention reaches. The piece quotes a researcher noting how embedded those tendencies are, which cuts both ways: edits can fail to remove bias, and they can also bend a model toward a different set of assumptions.
Chinese open models trail the top closed systems by roughly 3 percent on benchmarks while taking the largest share of Hugging Face downloads. Onboarding them is already routine; deciding whose values ship alongside them is the newer problem.
