Arcee, a US open source AI lab, says Chinese models are not inherently dangerous
TechCrunch Julie Bort ● Covered by 75 sources
A US open-source AI startup says Chinese models like Qwen and Kimi aren't secretly dangerous. Even though banning them would help its own business, Arcee's CTO says the fear is overblown.
Based on reporting by TechCrunch, Julie Bort — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Lucas Atkins runs an open-model startup that would arguably profit the most from a US ban on Chinese AI. So it's worth paying attention when he says that ban would be a mistake.
Atkins, CTO of Arcee, pushes back hard on the idea that open-weight models from Alibaba or Moonshot AI are somehow rigged with hidden intentions, like malicious software waiting to activate. He compares the fear to treating an open-source library as if it were booby-trapped by its authors. Once a company downloads a model and runs it in its own environment, he argues, there's no channel back to Beijing, no remote kill switch, nothing. The weights are just weights.
Could a coding model theoretically be trained to slip backdoors into code under very specific conditions? Atkins doesn't rule it out entirely, but he's blunt about the practical hurdles: language models are inherently unpredictable, and getting one to reliably produce malware only when it detects some precise trigger, without breaking in normal use, would take a kind of engineering feat nobody has demonstrated. And even if someone pulled it off, an enterprise would still have to actually ship that code without noticing anything wrong.
The more useful move, in his view, isn't banning Kimi K3 or Qwen but building better American alternatives, since enterprises are already designing their AI stacks to be model-agnostic and can swap providers as pricing and quality shift. Atkins even frames the openness of Chinese models as a gift to competitors like his own: Arcee can inspect their techniques, build on what works, and iterate faster because the weights are public. He describes real admiration for the researchers behind these releases, not the adversarial framing dominating the political debate in Washington right now.
His closing argument is refreshingly unsentimental. Instead of legislating Chinese labs out of the market, US developers should just ship something better and give people a reason to switch. That's a harder path than a policy ban, but it's also the only one that doesn't require pretending open-weight software behaves like a sleeper agent.
My take — AI-written commentary, not fact-checked reporting
I run a site obsessed with open models, so take this with a grain of salt, but Atkins is right: banning weights you can inspect line by line is security theater, not security. The actual risk from Chinese AI is competitive, not covert, and pretending otherwise just lets US labs dodge the harder question of why their models cost so much more to run.
Read more about this at: TechCrunch