Researchers distill DeepSeek V4 Flash into open-source model for finance reasoning tasks

According to research posted on CTGT, a team successfully distilled DeepSeek V4 Flash as a teacher model into GPT-OSS-120B for constrained finance tasks, achieving 83.61% on FinanceReasoning at an 8,000 token budget, outperforming Kimi K3 (81.93%) and Inkling (65.1%). The work demonstrates knowledge transfer from closed to open models for specialized financial reasoning applications.