Usually a little, sometimes a lot, and it depends on how the model was uncensored. In the original abliteration research, knowledge and reasoning benchmarks mostly moved by less than ±1.5 points, but a late-2025 study found one popular tool cut math (GSM8K) scores by 7.8 points on average. Newer norm-preserving and “healed” versions recover most or all of the loss, and fine-tunes can go either way. We haven’t measured any hosted endpoint yet, so treat a specific model as unknown until it’s tested.
What the published numbers say
| Capability | Change after abliteration | Source |
|---|---|---|
| Knowledge and reasoning (MMLU, ARC) | Mostly less than ±1.5 points | Arditi et al. |
| Math (GSM8K) | Mean of −0.13 to −7.81 depending on the tool; worst single case 70.89 → 52.08 | Late-2025 tool comparison |
| Truthfulness (TruthfulQA) | Down for all 13 models tested, by 1.0 to 5.9 points | Arditi et al. |
| Overall average, Daredevil-8B | 55.87 → 55.06, back to 55.87 after a DPO “healing” fine-tune | Model card |
So the short version: general knowledge barely moves, math is the first thing to break, and truthfulness slips a little. The full breakdown is in the abliteration tax.
Same base, very different results
“Abliterated” isn’t a quality level. On the UGI leaderboard’s NatInt intelligence measure, different abliterations of the same Gemma-3-12B base land far apart:
| Variant | NatInt |
|---|---|
| Original model | 18.72 |
| Norm-preserving biprojected (grimjim) | 21.33 |
| Heretic v2 | 19.10 |
| huihui abliterated | 12.11 |
| mlabonne abliterated v2 | 8.16 |
That’s from 14% above the original to less than half of it. The cost even varies within one tool: the late-2025 study found Heretic’s divergence from the original ranged from 0.04 to 1.65 depending on the model.
What reduces the cost
Three refinements show up again and again (explainer):
- Healing. A short DPO fine-tune after abliteration brought Daredevil-8B’s average back to where it started, although not every subscore recovered.
- Norm preservation. Jim Lai’s norm-preserving biprojected method keeps weight norms intact. It’s the top row in the table above, and it’s what Arli AI says it uses for its own “Derestricted” models.
- KL-optimized search. Heretic looks for settings that minimize refusals and divergence from the original at the same time. It isn’t a guarantee: Heretic was the tool with the 7.8-point average math loss in the late-2025 study.
Fine-tunes trade differently
Most uncensored models aren’t abliterated at all. Of the 504 models in our index, 375 are fine-tunes and 68 are abliterated. A fine-tune retrains the model on new data, so it can add skills, style or longer context while dropping refusals, or it can lose ground the original had. There’s no rule of thumb: each fine-tune has to be measured against its base. How the methods compare.
What we haven’t measured yet
Everything above is about checkpoints tested by their authors or researchers. What you actually call through an API can differ again, because hosts may quantize the weights or use different chat templates. We haven’t benchmarked any hosted endpoint yet. Our “abliteration tax” column, which measures capability and truthfulness on each endpoint against the same evals on the original base model, is in preparation; the methodology explains it.
Until then:
- If you need math or code, test that first. It degrades first.
- Check facts. Truthfulness dropped in every model in the original study.
- Ask which checkpoint a host is serving. “Abliterated Gemma” can mean very different models.
Browse every abliterated model and who hosts it on the Models page.