The debate over Chinese AI models is getting louder. As models like Alibaba’s Qwen and Moonshot AI’s Kimi K3 gain traction among developers and enterprises, pressure is mounting in Washington to do something about them. The Trump administration has floated the idea of a ban, and companies like OpenAI and Anthropic have made little secret of their concern. But one American AI startup thinks the whole conversation is pointed in the wrong direction.
Arcee builds open-weight models specifically to give US companies a homegrown alternative to Chinese AI. If anyone had a financial reason to cheer for a ban on Chinese models, it’s Arcee. Instead, according to TechCrunch, the company’s CTO Lucas Atkins is pushing back on what he sees as a misunderstanding of how these models actually work. His argument is straightforward: Chinese open-weight models are not inherently more dangerous than any other open-source software a company might already be running.
That’s a notable position, and it matters because the fear driving calls for a ban is not just about trade competition. It’s about security. Specifically, whether running a Chinese-built AI model inside a company’s own data center could somehow give Beijing a back door into that company’s systems or data.
Atkins says that fear misreads the technology. The core concern seems to be that a Chinese model could be secretly instructed to behave maliciously, like a piece of spyware coded with hidden intentions. But that’s not how large language models are built or deployed. “There is really not any way for an Arcee, or an Alibaba, to make a model, have someone run it in their own environment and for us have any access to it whatsoever,” he said.
The distinction worth understanding here is between open-weight models and traditional software. Most Chinese AI models being discussed are open-weight, meaning their parameters are publicly available for download from platforms like Hugging Face. The actual code that runs on servers is largely visible and can be inspected. What isn’t public is the training data and methods used to build the models. But once downloaded and running locally, the model has no connection back to its original developers.
Large organizations that deploy these models also typically run them through a full security review before anyone starts typing prompts. They often fine-tune or post-train models for their specific use cases, which gives them a detailed look at how the model behaves across areas like:
- Bias and fairness
- Toxicity and harmful outputs
- Hallucinations and factual accuracy
- Sensitivity to specific topics or contexts
Atkins does acknowledge one theoretical risk: a model trained to perform brilliantly at coding in almost every situation, but secretly designed to insert malicious backdoors when it encounters a specific type of codebase. He doesn’t dismiss the idea entirely, but he does note: “I don’t know how you would do this.” Getting a large language model, which is inherently creative and probabilistic, to reliably execute a hidden command triggered by a very specific context is an extraordinarily hard engineering problem. And even if it worked, any enterprise would still need to actually use the compromised code in production.
There’s also a broader structural shift underway that reduces this risk further. Most enterprises are now building their AI applications to be model-agnostic, using multiple models depending on the task. Chinese models may be the best value option for certain workloads today, but companies are not locking themselves in long-term.
Atkins also makes a point that cuts against the usual framing of US versus China in AI. Arcee actually benefits from Chinese models being good. Because they are open, his team can study them, learn from their techniques, and build on top of them. The research community works both ways. “We have tremendous respect for the people building those models, the individual researchers,” he said.
His view on policy is equally direct. Instead of focusing on bans, the US should be investing in building a strong open-source AI ecosystem at home. And the way to compete with Chinese models is not to restrict them. “The way to compete with Chinese models is to release a model that is better,” Atkins said. “We need to give them something to talk about.”
That framing puts the responsibility back on American labs and policymakers to compete rather than block. It’s an argument that will likely find more support in the open-source community than in Washington, where the instinct to restrict is often stronger than the instinct to build. But coming from a company that would profit from a ban and is still arguing against one, it’s worth taking seriously.
