Start-ups are modifying Chinese open-weight AI models to address censorship, security and other concerns from Western customers. The bigger question is whether a model can be adapted to a new market without simply replacing one set of values with another.
Watch Desk analysis
What happened
Rest of World reports that AI safety start-up Hirundo edits model weights to change geopolitical responses, improve security and reduce what its chief executive calls dangerous knowledge. Its co-founder Ben Luria says the aim is not to make models anti-Chinese, but to make them suitable for the expectations of the companies deploying them.
Boston-based Lazarus AI has also adapted Alibaba’s Qwen models, the report says. The company identified topics it said were subject to censorship, generated related prompts and used answers from Google’s Gemini Flash in its process. Former chief scientist Eric Hartford says its ReAligned models are deployed at about a dozen US companies. Those are company claims reported by Rest of World, not independent assessments of the models’ safety or effectiveness.
The Rest of World report also describes users who see little need to alter Chinese models, and researchers who argue that alignment depends on the cultural context and application. There is no single neutral setting waiting to be found under the bonnet.
Why it matters
Open-weight models can be adapted and run in ways that closed services may not allow. That makes them attractive to organisations seeking lower costs or more control, but raises practical questions about security, censorship and whose expectations shape a model’s answers.
The work described here is not just removing refusals. The companies say they are changing model behaviour for particular customers and uses. Whether those changes hold up across different prompts and real deployments matters more than a reassuring label such as “realigned”.
Our read
This is a real commercial response to a real deployment problem, but “Western values” is not a technical specification, and safety is not a finish line. Customers should ask what was changed, how it was tested and which risks the intervention is meant to address. Model editing can make a system more useful for a particular context; it cannot settle who gets to define the context.
What to watch
- Whether Hirundo and Lazarus publish clearer testing methods and independent evaluations.
- Whether their customers disclose which models and use cases they have adapted.
- How model behaviour changes across languages, political topics and security-sensitive tasks.
Discussion spark: When companies adapt an AI model for a local market, who should set the standard: the customer, the model maker, or the communities represented in its answers?
Sources and evidence
- Startups recalibrate Chinese AI for the West – Rest of World (2 November 2026, 11:01 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.