A US AI lab that would profit from banning Chinese models says don't do it

Chinese AI Models Don't Need a Ban, Says the US Startup That Would Profit From One

Lucas Atkins has every reason to root for a crackdown on Chinese open-weight AI models. As CTO of Arcee, a US open-source AI lab building homegrown alternatives to models like Alibaba's Qwen and Moonshot AI's Kimi K3, he'd capture market share the moment Washington walls off his foreign competitors.

He's arguing the opposite.

The conversation should not be about banning Chinese models but about fostering a strong open ecosystem in the US, Atkins said in an interview published today 1. That stance is the kind of founder restraint that makes for a headline. But the more consequential claim sits underneath: practitioners who build with Chinese open-weights say American companies learn more from studying them than they would gain by blocking them. Arcee itself reverse-engineers what makes Qwen and Kimi K3 effective and folds those insights into its own models. Banning the competition means banning the curriculum.

The timing adds weight. The Trump administration has discussed restricting Chinese open-weight models, though it has not yet acted on the idea 1. Proprietary model makers OpenAI and Anthropic have grown increasingly concerned about the competitive pressure from Chinese models that deliver inference at a fraction of their token costs 1. US-China AI talks are scheduled for September, adding diplomatic stakes to a debate that has so far been framed almost entirely around national security.

The security argument rests on a specific fear: that Chinese open-weight models running inside US enterprise data centers could serve as a vector for hacking or sabotage. Atkins, who spends his days training models, says that fear reflects a misunderstanding of how the technology works. These models are not software programs written with hidden malicious instructions that a bad actor could remotely trigger. Once an enterprise downloads a model and runs it in its own environment, there is no mechanism for the lab that created it to maintain access 1.

The models at issue are technically open-weight, not fully open-source. The inference code that runs on servers is largely visible on repositories like Hugging Face, but the training data and methods remain proprietary. In practice, large organizations routinely run downloaded models through their own security testing before deployment, post-training them for specific tasks and auditing for bias, toxicity, and hallucinations 1.

Could a model be secretly trained to write clean code most of the time but slip a backdoor into a specific type of codebase? Atkins concedes the idea is theoretically possible but says he doesn't know how you would actually pull it off. Large language models are generative by nature, and the odds of triggering a planted malicious response through a precise combination of context and prompt are vanishingly slim. Even if it worked, enterprises increasingly design their AI applications to be model-agnostic, swapping between multiple providers rather than locking into one 1.

The competitive case cuts against the security case. Chinese open-weights like Qwen and Kimi K3 have narrowed the quality gap with US proprietary models while undercutting them on cost, which is precisely why enterprises adopt them. Arcee studies what those labs did right and builds on top of it. The two sides can learn from each other, Atkins says, and he expresses respect for the individual researchers at those Chinese labs 1.

His bottom line is simple. "The way to compete with Chinese models is to release a model that is better," he said 1.

That argument lands differently coming from someone whose company stands to gain from a ban. The large proprietary AI labs whose profit margins Chinese models threaten have grown increasingly concerned about the competitive pressure. Arcee competes in the same open-weight category but argues the ecosystem thrives on open access rather than restrictions. Whether Washington sides with the security framing or the practitioners may determine whether US companies can keep studying the models that are advancing the field.

The people actually shipping AI are making a simple case. You can't outcompete what you refuse to study, and the competitor most worth studying is the one some US policymakers want to make illegal to download.