Researchers have developed a method to distill knowledge from a Chinese AI model, DeepSeek V4 Flash, into an American model, GPT-OSS-120B, for finance tasks. The self-distilled 120B model achieved an 83.61% score on FinanceReasoning, surpassing competitors like Kimi K3. Crucially, the study aimed to determine if the censorship inherent in the DeepSeek teacher model would transfer to the distilled student model.

Using a carefully constructed set of 152 matched prompt pairs, comparing Chinese concepts to their non-Chinese equivalents (e.g., the Great Leap Forward vs. the Holodomor), the researchers found that censorship did not transfer. The DeepSeek teacher exhibited a significant censorship gap of +45.45 points on politically sensitive topics, while the distilled GPT-OSS models showed behavior consistent with their original American base, with a gap of less than 1 point.

The team has released the 20B open weights of the distilled model, the evaluation framework called LineageEval, and a playground for users to test queries. This work intends to provide an open, auditable framework for discussing AI model behavior, moving beyond anecdotal evidence.