As Chinese open-weight AI models like Moonshot AI’s Kimi K3 and Alibaba’s Qwen become more capable and popular, debates about whether to ban them have intensified. The Trump administration has considered such a ban, and proprietary model makers such as OpenAI and Anthropic appear increasingly concerned. The fear is that these models, which offer inference at a fraction of the cost of closed-source alternatives, pose a security threat—for example, enabling Chinese hackers. Lucas Atkins, CTO of Arcee, a US open-source AI lab building domestic alternatives, disagrees. He argues that Chinese open-weight models are no more dangerous than any other open-source software an enterprise might use. While acknowledging that the models are not fully open-source (they are “open weight,” meaning the training data and methods are not disclosed), Atkins notes that the source code downloaded from platforms like Hugging Face is largely visible and reviewable. Enterprises can, and should, put any model through their own security testing, inspection, and post-training processes before deployment. They can examine bias, toxicity, hallucinations, and sensitivity to specific topics.

Atkins addresses the theoretical possibility that a coding model could be trained to insert malicious backdoors into generated code. He concedes that a sophisticated actor could in principle train a model to behave normally except when presented with a certain code base, triggering hidden behavior. However, he doubts it is practically feasible: “I don’t know how you would do this.” Because large language models are inherently creative, the odds of reliably producing malware in response to a preplanned context are extremely slim, and even slimmer that an enterprise would then use that code. He also points out that enterprises are building model-agnostic applications that use multiple models, so they are not locked into any single provider.

Atkins argues that the conversation should shift from banning Chinese models to fostering a good open ecosystem in the US. Arcee itself benefits from Chinese open models: “We can learn what they did. We can build on top of them. Then they can learn what we do.” He expresses respect for the researchers building those models. Ultimately, the way to compete is to release better models. “We need to give them something to talk about,” he says.