GPT-OSS-120B Achieves 83.61% on FinanceReasoning via Distillation
July 30, 2026
Distilling DeepSeek V4 Flash into GPT-OSS-120B yields a model that scores 83.61% on FinanceReasoning without transferring the teacher's censorship. The 20B open-weight model outperforms Kimi K3 and Inkling within a constrained 8k token budget.
HOW THIS AFFECTS YOU
●
builderYou can use high-performing closed models to train smaller, uncensored open models.
●
researcherDistillation can successfully decouple reasoning capabilities from alignment/censorship constraints.