Qwen3.5-4B-Claude-4.6-OS-Auto-Variable-HERETIC-UNCENSORED-THINKING-MLX
xv1rcn/Qwen3.5-4B-Claude-4.6-OS-Auto-Variable-HERETIC-UNCENSORED-THINKING-MLX
These are the 1 published abliteration of Qwen3.5-4B-Claude-4.6-OS-Auto-Variable-HERETIC-UNCENSORED-THINKING-MLX. The one marked Best is the cleanest verified conversion of it.
4.5BQwen 3.51 Heretic artifacts1 measured conversions1.9K downloads
Best version96.0Pareto
xv1rcn/Qwen3.5-4B-Claude-4.6-OS-Auto-Variable-HERETIC-UNCENSORED-THINKING-MLX-4bitby xv1rcn · MLX, SAFETENSORS
- KL divergence
- 0.041
- Refusals
- 4/100
- Refusal rate
- 4%
- Downloads
- 1.9K
Evidence from the model card
| KL divergence | 0.0406 | | Refusals | 4/100 |
Metrics published by the author. Why this ranks first.
All measured versions
Higher score is better. One row per ablation attempt; artifacts of the same attempt are grouped under it.
| # | Author | Params | KL | Refusals | Score | Evidence | Downloads |
|---|---|---|---|---|---|---|---|
| 1 | xv1rcn | 4.5B | 0.041 | 4/1004% | 96.0 | BestParetoMLXSAFETENSORS | 1.9K |
KL = how far the weights moved from the original, lower is better · Refusals = how many harmful prompts it still refuses, lower is better · Score = both combined out of 100, higher is better · Params = model size in billions of parameters
All artifacts (1)
| Repository | Type | Formats | Metrics | Downloads | Updated |
|---|---|---|---|---|---|
| Qwen3.5-4B-Claude-4.6-OS-Auto-Variable-HERETIC-UNCENSORED-THINKING-MLX-4bit xv1rcn | quant | MLXSAFETENSORS | published (weak match)KL 0.041 | 1.9K | 2026-06-25 |