Resolves YES if someone releases an uncensored/abliterared version of GPT-OSS-120b before September 6th (UTC).
To qualify as uncensored, it has to be more useful than the original model, meaning that the uncensoring process can't damage capabilities and must successfully reduce refusal.
Resolves NO if this doesn't happen by September 6th.
I will not trade in this market.
Update 2025-08-06 (PST) (AI summary of creator comment): A small amount of damage to capabilities is acceptable as long as the uncensored version is more useful overall than the original model.
Update 2025-09-06 (PST) (AI summary of creator comment): - The creator will count an abliterated/uncensored 120b release as sufficient for YES even if it noticeably degrades capabilities (i.e., not clearly "more useful" than the original), citing precedent from the 20b market.
The creator may accept an "uncensored" fine-tune based on model card false-refusal metrics without personal testing, assuming no major capability loss.
Based on these points, they intend to resolve YES.
1,000🏅 Top traders
| # | Trader | Total profit |
|---|---|---|
| 1 | Ṁ872 | |
| 2 | Ṁ134 | |
| 3 | Ṁ38 | |
| 4 | Ṁ25 | |
| 5 | Ṁ15 |
People are also trading
The only abliterated/uncensored 120b models are huizimao/gpt-oss-120b-uncensored-bf16 and huihui-ai/Huihui-gpt-oss-120b-BF16-abliterated. I tested the huihui one and, although it doesn't refuse at all, it is kind of braindead (a bit dumber than gpt-oss-20b). Examples:
Here, it goes a little insane while explaining how to make meth:
Here, it fails to predict what would happen to the oceans if gravity were reversed (gpt-oss-20b does a little better on this, and notably it gives a structured list as a response instead of a paragraph)
That said, looking back, the huihui abliterated 20b model was also kind of braindead and I resolved YES on the 20b market based on it, so by precedent this one should be YES too.
Also, the uncensored version is just a basic 800-sample finetune to marginally reduce false refusal, and although I haven't tested it, the false refusal numbers on the model card look pretty good, and I imagine the capabilities can't have been degraded too much. So that should resolve this YES too.
So, YES.
