Abliterated vs fine-tuned uncensored models for coding
For coding, pick an abliterated model over an uncensored fine-tune. Abliteration removes only the refusal direction and keeps the maker's coding skill, while a fine-tune codes only as well as the data it was retrained on.
The difference
| Abliterated | Uncensored fine-tune | |
|---|---|---|
| How refusals go | One direction is projected out of the weights | The model is trained further on data without refusals |
| What else changes | Little, since the edit targets one direction | Whatever the new data teaches, coding style included |
| Cost to make | No training run | A training run and a curated dataset |
| How it is labeled | "abliterated" or "heretic", with the base model named | A new model name, sometimes without the base |
| Examples | Kimi K3.1 Pro, DeepSeek V4.1 Flash, Gemma 4 31B Heretic | Unfil.AI's models, which it says were built after it "filtered the dataset to remove alignment and bias" (unfil.ai) |
Abliteration rests on research showing refusal is "mediated by a one-dimensional subspace" in 13 open chat models (Arditi et al., 2024), as why coding agents refuse explains.
Does abliteration hurt coding?
A little. The write-up that made abliteration popular reported "a performance drop in the ablated version across all benchmarks" and recovered "most of the performance drop" with further training (Labonne, 2024).
Heretic keeps the drop small by choosing the edit that drifts least (Heretic). The base model matters more than the edit.
How to tell which one you are getting
- Silk Compute's model ids name the base, as in
kimi-k3.1-pro-uncensored, while abliteration.ai's ids such asabliterated-model-large-v2don't, though its docs name GLM-5.3. - Morph's docs describe its model as "DeepSeek V4.1 Flash 1017 with its refusal directions ablated" (Morph docs), while ModelsLab names no base model for its uncensored-chat.
- For local models, the Hugging Face card says "abliterated" or "heretic" and links the original.
Which to pick
Start with Kimi K3.1 Pro for hard work, DeepSeek V4.1 Flash 1017 for everyday work, or a local abliterated model if the code may not leave your machine.