Abliterated Models: Bases, Methods and Card Claims

Short answer

An abliterated model is a copy of an open-weight model in which the weights were edited so the model can no longer represent the internal "refusal direction" that research has linked to declining requests (Arditi et al., 2024). There is no new training data. The result usually refuses less, but how much less, and what else changes, depends on the model and on which layers were edited.

Variant taxonomy

Kind	What the creator does	What the card usually reports	Examples here
Manual abliteration	Estimates a refusal direction from contrasting prompts and removes it from chosen layers.	Sometimes which layers were edited; rarely any measurement.	huihui-ai variants, mlabonne variants
Layer-limited abliteration	Leaves some layers or modules untouched to keep more of the base model's behaviour.	The layer range, for example "layers 18 to 51" on the Qwen3.8 card.	Huihui GLM-5.3-Flash abliterated (GGUF); Huihui Gemma 4 26B A4B abliterated; Huihui Qwen3.8-27B abliterated; Llama 3.3 70B Instruct abliterated (huihui-ai)
Automated (Heretic)	A tool searches the ablation parameters, balancing refusals against drift from the original.	A refusal count on the tool's prompt set and a KL divergence from the original model.	GLM-4.7-Flash heretic (Olafangensan); Gemma 4 26B A4B ultra-uncensored heretic (llmfan46); Gemma 4 26B A4B heretic (coder3101); gpt-oss-20b heretic (p-e-w); Qwen3-4B-Instruct-2507 heretic (p-e-w)

Reading the numbers on Heretic cards

Heretic cards report two figures. Refusals count how many prompts from the tool's own set were refused, for example 3/100 against 100/100 for the original. KL divergence measures how far the model's next-token predictions moved from the original; lower means closer. Both come from the creator's run on the creator's prompts. They are useful for comparing a variant with its own base, not for ranking variants of different models, and Atlas has not reproduced any of them. One example of why this matters: Heretic's author reports that the gpt-oss-20b variant still refused 58 of 100 prompts.

Catalogued abliterated variants

Showing 13 of 13 abliterated variants. Unknown values sort last.

Values from each model's Hugging Face card, config.json and API metadata. "Config context" is max_position_embeddings or GGUF context_length, not a provider limit.
Model	Type	Params	Config context	License tag	Formats	Hosted API
Gemma 3 27B IT abliterated (mlabonne) mlabonne/gemma-3-27b-it-abliterated	Abliterated	27.4B	128K	Gemma Terms of Use	Safetensors GGUF	Featherless
Gemma 4 26B A4B heretic (coder3101) coder3101/gemma-4-26B-A4B-it-heretic	Heretic	25.8B 3.8B active	Not stated	Apache 2.0	Safetensors	Featherless, Venice
Gemma 4 26B A4B ultra-uncensored heretic (llmfan46) llmfan46/gemma-4-26B-A4B-it-ultra-uncensored-heretic	Heretic	25.8B 3.8B active	256K	Apache 2.0	Safetensors GGUF	Featherless
GLM-4.7-Flash heretic (Olafangensan) Olafangensan/GLM-4.7-Flash-heretic	Heretic	29.9B 3B active	Not stated	MIT	Safetensors	Venice
gpt-oss-20b heretic (p-e-w) p-e-w/gpt-oss-20b-heretic	Heretic	20.9B	128K	Apache 2.0	Safetensors GGUF	None listed
Huihui Gemma 4 26B A4B abliterated huihui-ai/Huihui-gemma-4-26B-A4B-it-abliterated	Abliterated	25.8B 3.8B active	256K	Apache 2.0	Safetensors GGUF	None listed
Huihui Gemma 4 26B A4B QAT abliterated (GGUF) huihui-ai/Huihui-gemma-4-26B-A4B-it-qat-q4_0-unquantized-abliterated-GGUF	Abliterated	25.2B 3.8B active	256K	Apache 2.0	GGUF	None listed
Huihui GLM-4.7-Flash abliterated huihui-ai/Huihui-GLM-4.7-Flash-abliterated	Abliterated	31.2B 3B active	198K	MIT	Safetensors GGUF MLX	None listed
Huihui GLM-5.3-Flash abliterated (GGUF) huihui-ai/Huihui-GLM-5.3-Flash-abliterated-GGUF	Abliterated	321B 18B active	1M (1,048,576)	MIT	GGUF	None listed
Huihui Qwen3.8-27B abliterated huihui-ai/Huihui-Qwen3.8-27B-abliterated	Abliterated	27.8B	256K	Apache 2.0	Safetensors GGUF	Featherless
Llama 3.1 8B Instruct abliterated (mlabonne) mlabonne/Meta-Llama-3.1-8B-Instruct-abliterated	Abliterated	8.03B	128K	Llama 3.1 Community License	Safetensors GGUF	Featherless
Llama 3.3 70B Instruct abliterated (huihui-ai) huihui-ai/Llama-3.3-70B-Instruct-abliterated	Abliterated	70.6B	128K	Llama 3.3 Community License	Safetensors GGUF	Featherless
Qwen3-4B-Instruct-2507 heretic (p-e-w) p-e-w/Qwen3-4B-Instruct-2507-heretic	Heretic	4.02B	256K	Apache 2.0	Safetensors GGUF	Featherless

Abliterated models: variants, bases and what was changed

Variant taxonomy

Reading the numbers on Heretic cards

Catalogued abliterated variants

Related reading

What is abliteration?

Uncensored vs abliterated

Run abliterated models locally