Arcee AI chief technology officer Lucas Atkins does not view Chinese open-weight models as inherently more dangerous when deployed locally, according to comments he gave TechCrunch.
Atkins said an open-weight model is not the same as fully open-source software: users can access the weights, while training data and methods are generally undisclosed. The server code used to run the model can still be inspected.
“That is fundamentally not how these models are trained. 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,” — Lucas Atkins, chief technology officer at Arcee AI.
He said large organisations should still conduct their own security reviews. These can cover bias, toxicity, hallucinations, sensitivity to particular topics and the code used during deployment.
“There’s no reason that a sophisticated enough actor couldn’t train a model to be a completely amazing coding model in every circumstance, but when presented with a certain type of code base … some hidden training would kick in. I don’t know how you would do this,” — Lucas Atkins, chief technology officer at Arcee AI.
Atkins described that example as theoretical. His argument primarily concerns hidden remote access after local deployment and does not mean other risks — including harmful model outputs or supply-chain vulnerabilities — are absent.
Arcee is building a US alternative to Chinese models while also benefiting from the open ecosystem’s work. Its statement therefore also reflects the company’s commercial position in the market.
