A paid market for "realigned" Chinese open-weight models lets Western enterprises keep Qwen's low costs while answering censorship and security objections, which widens Alibaba's reach outside China without its involvement.
A small group of Western startups is selling edited versions of Chinese open-weight AI models to companies that want the models’ low cost without the censorship built into them. Rest of World reported on November 2 that Israel-founded Hirundo and Boston-based Lazarus AI are altering Alibaba’s Qwen models to strip out answers that follow Beijing’s political line, and in some cases to reduce security risks.
Lazarus AI’s realigned Qwen models are deployed at about a dozen US companies, according to Eric Hartford, its former chief scientist. He told Rest of World that his aim was to make the models "palatable" for enterprises and highly regulated industries. Lazarus serves the US government and regulated sectors with AI that runs on customers’ own servers.
The business exists because Chinese models have become hard for cost-conscious companies to ignore. Qwen has been downloaded more than 3 billion times, CBS News reported in October, and Chinese open-weight models rose from under 2% of global usage on the OpenRouter platform in late 2024 to more than 45% by June 2026.
Two methods for removing the same restrictions
The two companies work differently. Hirundo, co-founded by chief executive Ben Luria, edits a model’s internal weights directly. In tests reported by CBS News, Hirundo ran 500 prompts across 15 topics and found that the original Qwen produced censorship, propaganda-aligned framing or political bias in 89.8% of answers to sensitive political prompts. Its edited version did so in 2.8% of cases, while keeping reasoning, coding and math abilities intact, according to the company. Hirundo’s figures come from its own benchmark.
Lazarus AI took a different route, according to Rest of World. It identified banned topics from leaked censorship guidelines, generated prompts on those topics, and used answers from Google’s Gemini Flash model to realign Qwen. Afterward, the models supplied information that had been missing from their answers, suggesting the knowledge had been learned during training and then suppressed.
Luria said many simpler tools that remove a model’s refusal behavior leave censorship and misinformation in place. "We’re not making the model anti-Chinese," he told Rest of World, describing the goal as bringing models up to standard on safety.
Security worries strengthen the commercial case
Censorship of topics such as the 1989 Tiananmen crackdown is the most visible issue, but the commercial case rests partly on security. CrowdStrike research cited by Rest of World found that DeepSeek’s R1 model was up to 50% more likely to produce code with severe security vulnerabilities when prompts touched politically sensitive topics. A Booz Allen study cited by CBS News found that Qwen-generated code had 130% more vulnerabilities when the project was described as being for the US government.
Policy pressure is rising at the same time. The Trump administration barred civilian government agencies from using DeepSeek products in October, Rest of World noted. Airbnb, Uber and legal AI company Harvey have built their own models on top of Chinese ones, and Europe’s OpenEuroLLM sovereign model is being developed on Qwen, according to the publication. EastFrontier reported in July that US companies were routing 46% of AI tokens through Chinese models.
Not everyone sees realignment as neutral. Oriane Peter, a researcher at King’s College London, called aligning a model to universal Western values "an impossible standard" and warned that centralized control over alignment in Silicon Valley could make models more alike. Graham Webster of Stanford’s DigiChina Project said the stakes depend on use: a product used in Taiwan matters far more than a coffee-shop website builder. Alibaba did not respond to CBS News’s request for comment.
US procurement rules will decide how big this market gets
The size of the market depends on how Washington treats models that are Chinese in origin but altered in the West. A federal rule that bans products built on Chinese models, edited or not, would close the US government channel that Lazarus serves; guidance that accepts independently audited versions would open it wider.
A second signal will come from testing. Hirundo’s results so far are its own; independent comparisons of edited and original Qwen models on the same prompts would show whether realignment removes censorship as completely as its sellers claim.