Research · Hacker News ·
Show HN: Distilling DeepSeek into GPT-OSS doesn't transfer censorship. Try it
Researchers distilled DeepSeek V4 Flash into GPT-OSS-120B for finance tasks and released 20B open weights. In tests using 152 matched political prompts, the teacher showed a large China-specific response gap, while the distilled model retained behavior similar to its American base rather than transferring that pattern.