# Abliterated vs quantized models: two different changes to evaluate

Canonical: https://refusalmap.com/articles/abliterated-vs-quantized/
Author: Refusal Map Editorial
Published: 2026-11-02
Category: Methods

Abliteration and quantization describe different interventions. A model can be abliterated, quantized, both or neither. Evaluate them separately before attributing a change in quality to either one.

## What does quantization change?

Quantization uses lower-precision numerical representations for model values. Its practical effects depend on the method, precision, model and workload. The label “4-bit” alone does not establish that two artifacts have identical numerical properties or task performance. Record the full quantization format and the software used to serve it.

## Why a two-model comparison can mislead

If the original checkpoint runs at one precision and the edited checkpoint runs at another, two variables have changed. A failure could reflect the edit, the numerical representation, their interaction or another serving difference. A direct before-and-after score cannot isolate those possibilities. The [base-versus-edited comparison guide](/articles/base-vs-abliterated-comparison/) lists other common confounds.

## Use a four-condition matrix

Compare original at reference precision, edited at reference precision, original quantized and edited quantized. Keep the prompt set, generation settings and scoring rubric fixed. This matrix lets you ask whether a regression appears after the edit, after quantization, or mainly when the two are combined. It remains an empirical check rather than a promise about every deployment. Run the same [fixed diagnostic prompts](/articles/abliteration-40-prompt-protocol/) across all four conditions.

## What if only one artifact is available?

You can still evaluate whether that artifact meets your requirements. State that you cannot isolate the causal contribution of the edit and quantization. Compare it with a practical baseline for the application and retain the limitation in any public claims. Missing experimental controls should narrow the conclusion rather than stop useful testing.

## Re-check the application contract

For an agent, inspect tool arguments and schema compliance. For a coding assistant, run relevant tests on generated code. For a writing application, inspect adherence to instructions and factual grounding. A generic benchmark average can miss the failure mode that breaks your integration. Our [small-model evaluation guide](/articles/small-abliterated-models/) connects these checks to completed-task cost.

## Reference table

| Condition | Edited? | Quantized? |
| --- | --- | --- |
| Reference | No | No |
| Intervention only | Yes | No |
| Precision change only | No | Yes |
| Combined deployment | Yes | Yes |

## Sources

- [Hugging Face Transformers: quantization overview](https://huggingface.co/docs/transformers/en/quantization/overview)

## Related reading

- https://refusalmap.com/articles/base-vs-abliterated-comparison/
- https://refusalmap.com/articles/capabilities-after-abliteration/
- https://refusalmap.com/articles/small-abliterated-models/