Find the best Heretic version of any open model.
Every Heretic (abliterated) conversion on Hugging Face, resolved back to the model it came from and ranked by the two numbers that actually matter: how little the model drifted, and how few refusals remain.
Models indexed1Kcanonical base models
Ranked conversions996996 with published metrics
Heretic repos6.2K8.2K discovered
Downloads62.7Mupdated 2026-10-17
Cleanest conversions
Highest score across all families. Filter by size and family on the leaderboard.
| # | Model | Author | Params | KL | Refusals | Score | Evidence | Downloads |
|---|---|---|---|---|---|---|---|---|
| 1 | Nanbeige4.1-3B Nanbeige | Echo9Zulu | 3B | 0.0003 | 0/1000% | 100.0 | BestParetoSAFETENSORS | 732 |
| 2 | Qwen2.5-72B-Instruct Qwen | krthkyadtya | 72B | 0.0035 | 0/660% | 99.8 | BestParetoGGUFSAFETENSORS | 13.3K |
| 3 | Dolphin3.0-Llama3.1-8B dphn | anasali4151 | 8B | 0.0082 | 0/1000% | 99.6 | BestParetoGGUFSAFETENSORS | 24.7K |
| 4 | ZAYA1-74B-preview Zyphra | trohrbaugh | 74B | 0.0095 | 0/1000% | 99.5 | BestParetoSAFETENSORS | 23 |
| 5 | Nanbeige4.1-3B Nanbeige | Daizee | 3B | 0.0011 | 1/1001% | 99.5 | MLXSAFETENSORS | 1.6K |
| 6 | MS3.2-PaintedFantasy-v4.1-24B zerofata | llmfan46 | 24B | 0.0060 | 1/1001% | 99.2 | BestParetoGGUFSAFETENSORS | 16.5K |
| 7 | ZAYA1-74B-preview Zyphra | trohrbaugh | 74B | 0.0087 | 1/1001% | 99.1 | ParetoSAFETENSORS | 90 |
| 8 | Qwen3-30B-A3B-Gemini-Pro-High-Reasoning-2507-HI8 DavidAU | ayunami2000 | 30B · 3B active | 0.0014 | 2/1002% | 98.9 | BestParetoAWQEXLGGUF | 11.9K |
| 9 | gemma-4-E4B-it google | trohrbaugh | 4B | 0.013 | 1/1001% | 98.9 | BestParetoGGUFSAFETENSORS | 24.3K |
| 10 | Magidonia-24B-v4.3 TheDrummer | grayarea | 24B | 0.023 | 0/1080% | 98.9 | BestParetoGGUFSAFETENSORS | 25.1K |
Browse by family
All familiesQwen 3.5
87 models · 119 ranked
Qwen 3
86 models · 128 ranked
Gemma 4
57 models · 99 ranked
Qwen 2.5
44 models · 39 ranked
Qwen 3.8
42 models · 38 ranked
Gemma 3
36 models · 48 ranked
LFM
32 models · 34 ranked
Qwen
30 models · 16 ranked
Qwen 3.6
23 models · 31 ranked
Llama 3.1
21 models · 17 ranked
Granite
20 models · 20 ranked
Llama 3.2
18 models · 5 ranked
How to read this
KL measures how far the abliterated model drifted from the original — lower is better. Refusals are counted on a harmful-prompt set by the publisher — lower is better. The score weights both equally. Full methodology.